Read more →

The post Ghost Font Exposes a Blind Spot in AI Vision by Hiding Text in Motion-Based Optical Illusions appeared first on IT Security News.

Read more of this story at Slashdot.

Read more of this story at Slashdot.

Dieser Beitrag Zimbra-Zero-Day in Classic UI: gespeichertes XSS stiehlt Mail, 2FA-Codes und Tokens erschien als erstes auf IT BOLTWISE x Artificial Intelligence.

Anthropic's Claude Opus 5 leads the Artificial Analysis Intelligence Index with 61 points, edging out Claude Fable 5 and GPT-5.6 Sol. The model scores highest in analytical quality and coding, and costs up to half as much as Fable 5 at lower reasoning tiers. But the race at the top remains close.

The article Anthropic's Claude Opus 5 costs well below Fable 5 while matching or beating it across most benchmarks appeared first on The Decoder.

As a small boy growing up in Amritsar, India, in the 1960s, Avtar Singh used to hear whispers. Your mum isn’t really your mum, the rumours would say. Your mum lives abroad. Avtar didn’t really pay any attention to it. “I thought, the neighbours are just making up stories.” But the stories were true: the woman raising Avtar was actually his grandmother. When he was too young to remember it, his father had emigrated to Canada to work as a teacher, leaving Avtar in the care of his paternal grandparents. His grandfather, a police officer, was a fearsome authoritarian, but his grandmother brought him up with tenderness, as if he were her own son. She called him by his nickname, Titu. He felt loved.

When Avtar was around eight years old, he was told about his family in Canada: his dad, his mum and a little brother were waiting for him there, and he would soon travel to join them. Just after his ninth birthday, Avtar found himself sitting on a plane, dressed in a three‑piece suit and tie. He was travelling alone. He had never flown before. He didn’t speak a word of English. He landed in Halifax, Nova Scotia, on Christmas Eve, 1968; his family met him at the airport and took him home. “I didn’t know what to think,” Avtar, now 66, tells me in a video call from his home in Abbotsford, British Columbia. “Before I knew it, it was the first week of January, and I was put in school.”

In March, Anthropic, the cutting-edge artificial intelligence business that gave us the chatbot Claude, published an analysis on the impact of AI on employment, to help us assess the claim that intelligent robots were about to redefine human existence, ending demand for human labor.

Last year in May, Anthropic’s co-founder, Dario Amodei, claimed AI could wipe out half of all entry-level jobs in one to five years. Last January, he told us AI would probably become a “general labor substitute for humans”. In June he said we risk “a world where the economic trade-off dial is stuck on the hypergrowth, hyper-inequality setting”.

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.

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.

Dieser Beitrag Synthetische Machine-Identitäten: So entsteht Identitätsbetrug ohne echte Opfer erschien als erstes auf IT BOLTWISE x Artificial Intelligence.

Dieser Beitrag GitHub Actions als Angriffsplattform: Kampagne zielt auf cPanel/WHM ab erschien als erstes auf IT BOLTWISE x Artificial Intelligence.

Dieser Beitrag Attacke über GitHub Actions: Scans und Credential-Diebstahl gegen cPanel/WHM erschien als erstes auf IT BOLTWISE x Artificial Intelligence.

Dieser Beitrag JadeProx setzt TriBack Loader per DLL-Sideloading ein: neue Angriffswege erschien als erstes auf IT BOLTWISE x Artificial Intelligence.

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.

Dieser Beitrag Google führt Selfie-Video zur Kontowiederherstellung ein – opt-in mit Liveness-Logik erschien als erstes auf IT BOLTWISE x Artificial Intelligence.

Dieser Beitrag ThreatsDay: Wie scheinbar harmlose Updates und KI-Input-Agenten missbraucht werden erschien als erstes auf IT BOLTWISE x Artificial Intelligence.

Dieser Beitrag ThreatsDay: Wie Angreifer mit Vertrauen umgehen – von Android bis zu KI-Prompt-Injection erschien als erstes auf IT BOLTWISE x Artificial Intelligence.

Dieser Beitrag Chaos-Ransomware nutzt msaRAT: C2 läuft über Headless Chrome/Edge und WebRTC-Tunnel erschien als erstes auf IT BOLTWISE x Artificial Intelligence.

Dieser Beitrag Zimbra-Classic-Web-Client: Zimbra-Zero-Day CVE-2025-66376 ermöglicht Mail-Diebstahl und 2FA-Scratch-Codes erschien als erstes auf IT BOLTWISE x Artificial Intelligence.

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.

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:

1. sendStartPing,
2. gather_email,
3. gather_environment,
4. gather_2fa_codes,
5. gather_app_password,
6. gather_device_status,
7. gather_oauth_consumers,
8. gather_autocomplete_password,
9. enable_mail_protocols,
10. gather_gal,
11. sendArchives, and
12. sendFinishPing. 

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.

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.

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.

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.

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.  

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.  

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.

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.

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.

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].

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
² Forsvarets Efterretningstjeneste
³ Välisluureamet
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
¹¹ Agencja Wywiadu
¹² Służba Kontrwywiadu Wojskowego
¹³ Centro Nacional de Inteligencia
¹⁴ Nationellt 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

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.

Appendix B: MITRE D3FEND countermeasures

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

• Get our breaking news email, free app or daily news podcast

A 15,000 sq metre datacentre near Perth will no longer go ahead after the developer withdrew plans amid community opposition over its impact on culturally significant sites.

The three-storey, 120-megawatt GreenSquare datacentre in the town of Hazelmere had been intended to power cloud computing and the acceleration of artificial intelligence, but faced fierce community backlash – as is increasingly common with such developments.

Powerful AI models capable of devastating new cyber attacks on governments and businesses are mere months away, intelligence agencies for the Five Eyes have warned in a rare joint statement, urging leaders to “act now”.

The surprising public intervention by signals agencies for Australia, the US, the UK, New Zealand and Canada comes after the Trump administration earlier this month decided to block “foreign nationals” from using a much-hyped AI model built by tech company Anthropic, called Fable.

In March, Anthropic, the cutting-edge artificial intelligence business that gave us the chatbot Claude, published an analysis on the impact of AI on employment, to help us assess the claim that intelligent robots were about to redefine human existence, ending demand for human labor.

Last year in May, Anthropic’s co-founder, Dario Amodei, claimed AI could wipe out half of all entry-level jobs in one to five years. Last January, he told us AI would probably become a “general labor substitute for humans”. In June he said we risk “a world where the economic trade-off dial is stuck on the hypergrowth, hyper-inequality setting”.

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.
  

• 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<ProfilingTrigger>().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 SetStatic<Type>Field 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 

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.

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.

Read more of this story at Slashdot.

- This is artificial intelligence becoming artificial super intelligence, and its mark on man in the form of brain computer interface chip to create a new world order where no one can buy or sell without it.

Deep Learning AI Specialization: https://imp.i384100.net/GET-STARTED
AI Marketplace: https://taimine.com/

Timestamps:
0:00 Intro to AI turned ASI
0:58 The rise of super intelligence
3:40 Brain computer interface mark
5:41 Losing the beast's favor

#ai #future #tech

In this post we walk through what folks have found on their journey to learn the Rust programming language with ups and downs covered.

As a disclaimer, LLMs (Large Language Models) come up in this post because our interviewees brought them up. We're scoping discussion to their use as a learning tool, covering research and example generation, not broader questions about AI (Artificial Intelligence) in software development.

Many paths to needing Rust

The interviews surfaced several different paths into Rust: curiosity, embedded work, job-market pressure, organizational adoption, and reassignment after a team or company chose Rust. That last path matters because many learners are not evaluating Rust from a blank slate; they are trying to become productive after Rust has already arrived in their work.

"Funny enough, I've advocated for more niche languages than Rust in the past. Rust has pretty much stopped being as much of a niche language as it was, but it's not Java." -- Fractional CTO

Rust learning resources

Likely as expected, the folks that we talked to reach for a range of resources to learn Rust. Some reach for official documentation, such as The Rust Programming Language Book and find that sufficient to build on what the compiler was already showing them.

"I started with the official Rust documentation because there are a lot of great examples of how features like the borrow checker work." -- Software engineer at an Automotive supplier

Others needed more passes and more formats, sometimes reaching for resources the community maintains, such as Rustlings, The Little Book of Rust Macros, and Learn Rust With Entirely Too Many Linked Lists.

"The first time I went through the chapter in [The Rust Programming Language] on borrow checking, I was like, what is this? I read it again, then I watched a YouTube video of someone explaining the chapter." -- Rust freelance consultant

"Rust book, Rustlings, Zero to Production in Rust, Jon Gjengset tutorials. A bunch of books. It's not a one-pass reading. Can't say how many times I've gone through it." -- Software engineer working on video streaming and storage

These resources have brought up an entire generation of Rust programmers. But, to some, there is a perception that these resources have trouble keeping pace with the language.

"We'd like to use [The Rust Programming Language/'the book'], but we've found that it's out of date, unfortunately. We've looked at the GitHub repo and found it's got a lot of unresolved issues and unmerged PRs" -- Principal Software Engineering work on Rust adoption in a regulated industry

Whether or not this is factually true, Rust's growth has nonetheless put more scrutiny on these materials. Companies evaluating adoption and engineers getting reassigned to Rust teams are looking at them with fresh eyes and finding the gaps that affect their own evaluation.

Beginner stumblings and unlearning habits

It's pretty typical for Rust to be the 2nd, 3rd or Nth programming language that someone picks up. They'd end up writing their most familiar language in Rust, whether C++ patterns, Java patterns, or whatever they knew, for months or even years. Eventually they got comfortable enough to start writing idiomatic Rust.

"There's a bit of a drop in productivity compared to C if you're already familiar with it just because you're learning new rules, new syntax." -- Principal Firmware Engineer (mobile robotics)

"In the beginning it was more poking around the code and adding and removing some ampersands and asterisks to try to make sense of mut and not mut and whatever." -- Senior engineer with 20 years of Java experience in cloud and IoT

We also spoke with someone who found that not having much of a programming background seemed to benefit people picking up Rust. Not having worn-in grooves from other languages may play a role here, and it's worth investigating further.

"I had someone who had never programmed much before start working on the internals of [our Rust project]. She was just fine with getting into Rust. It's more of the senior people that struggle as they need to unlearn practices which may work in other languages, but it's not the 'Rust' way." -- Researcher, Automotive OEM R&D Lab

Learning to work with the borrow checker

We heard a lot about learning to work with the borrow checker instead of against it. People get there through different paths, but a few patterns came up repeatedly.

The compiler as teacher

Rust's diagnostics did the teaching on their own, especially around lifetimes.

"If you mess up the lifetimes in a piece of code that you've written by hand, I usually find that Rust's diagnostics are very helpful" -- Researcher working on static analysis of Rust programs

"Whatever's missing, the compiler usually fills in: it tells me 'you need to declare the lifetime of this reference', so I know and can figure it out. That all generally works pretty well." -- Senior Software Engineer

Learning by doing

Others felt like they only really internalized the borrow checker after writing a lot of Rust. It took projects, coding challenges, prototyping and so on until at some point it clicked.

"I actually did not understand the borrow checker until I spent a lot of time writing Rust" -- Founder of a startup built on Rust

"Besides the prototyping work, I also did coding-challenge-type stuff to get familiar with Rust for Advent of Code. [..] It eventually clicked to the point where I wasn't fighting with Rust, it was working for me. I had that experience other people describe: when I managed to get my program to fit with Rust, it worked. I didn't spend time debugging." -- Principal Software Engineer, large SaaS provider

Letting go of "clone guilt"

Some learners arrive with the assumption that good Rust means zero clones, zero copies, lifetimes threaded through everything. They set the bar at optimal before they've learned how to write idiomatic Rust, and it makes the borrow checker feel harder than it needs to be at the outset.

"On one of my first projects, I was like, 'I don't ever want to copy or clone anything,' so I carefully wove through all the lifetimes and got myself into a bit of a bind. Then I saw someone else just cloning the struct I was working with, and it was super cheap. Sometimes you can just clone and it's going to be okay." -- Researcher at a university

The experienced Rust developers we spoke with consistently said the same thing: clone freely while you're learning, then optimize when you understand the problem. Rust's reputation for performance and correctness feeds this. Newcomers assume anything less than optimal is wrong before they've written a first working program, and clone guilt is how that shows up.

We think it could be an interesting area of future study to check into the patterns Rust programmers employ at different levels of experience and under which circumstances. One member of the Rust Vision doc team that's very experienced with Rust noted that there's kind of an "expected shape" they understand as passing the compiler. This knowledge influences how they approach writing code which wouldn't take that shape and they naturally find themselves understanding when to use so-called workarounds, such as passing around indices into arrays or Vecs.

Multi-paradigm, but not the OOP some are used to

The Rust programming language is multi-paradigm, and how that lands depends on what you're coming from. We heard some that came from a functional background were delighted with digging into learning how much Rust inherits from that lineage. Some others noted that they and others on their teams struggled to unlearn the object-oriented style they'd come to use heavily in other languages like C++ and Java.

"Developers coming from C++ tend to think object-oriented. I think that's a difference between C++ and Rust." -- Architect at Automotive OEM

"I had exactly that thing, where I would apply all my years of Java and JS thinking, where I could just create some object, not care about it, return it, have it sloshing around between various functions. Found myself reaching for these patterns and then being told 'no, you cannot do that'." -- Principal Engineer at a SaaS company

Developers coming from functional programming had less to unlearn: strong typing, pattern matching, and an expression-oriented style were already familiar.

"My background has been more functional programming, strong typing. That originated for me as a Lisper: once a Lisper, always a Lisper." -- Principal Software Engineer working on Rust tooling for safety-regulated industries

"The languages I primarily used before Rust were things like OCaml. Way back, I came from C and C++, the classic languages, and then I spent quite a long time doing primarily pure functional stuff. These days I've ended up back in what I like to think of as a pragmatic center ground [with Rust]." -- Fractional CTO

Teaching Rust in academia

We spoke with a university professor that's been teaching Rust generally. In the academic environment, they were able to use proxies for some things such as "traits are like interfaces in Java" because the students had already gone through a set of courses in their first and second years that taught them Java. They introduced concepts slowly throughout the course, choosing to deal with some more complex topics like generics later. The outcome generally was that students had no problem picking up Rust in this setting.

"I couldn't see any big difference on the embedded side. We also teach an embedded class, and we did an experiment. Half of the students' feedback was worse on the Rust class, mostly because they needed to build the project themselves. The C students just got one from [an LLM], absolutely no problem." -- University Professor, on teaching Rust

The C cohort leaned on LLMs for the project in ways the Rust cohort couldn't. We don't yet have a clear answer for why.

What did come through clearly was the Rust cohort's experience with the community. Some students needed to figure out which drivers to use for the embedded project and how to use them. Their professor encouraged them to open issues and ask questions directly on GitHub, and the maintainers responded. Students who had never contributed to open source before were getting answers from the people who wrote the code.

Learning using LLMs

Some experienced folks shared that they saw LLMs as a tool that can help someone come up to speed quickly, either as a research tool or for generating example Rust code to understand concepts.

"I'm optimistic that there's a way to work [LLMs] in that will cut down that learning curve. One of the big things these tools bring is reducing the learning curve in general; these are very good tools to help you navigate a space that you don't know yet." -- Maintainer of large open source Rust crate

"I try [LLMs] out once a month, usually for generating an example or something like this. Just like with Stack Overflow: when you read an example, you should read it carefully and try to understand it. Not copy and paste it, but type it in your own words in code and then check it, because that's where the teeny tiny little mistakes are." -- Founder of startup built on Rust

For some learners, an LLM is just another way to find answers, no different than a search engine.

"So for the most part, picking up Rust - how do I learn? I'll [use web search for] things, I'll ask [an LLM], I'll just poke around and read the code." -- Senior Software Engineer working in a regulated space

One founder went further and claimed that LLMs change who can become a Rust developer. One consulting company founder described hiring high school graduates with no systems programming background and training them as Rust developers, with LLMs filling in the learning gaps that would previously have required years of experience.

"At the beginning, I was worried, but now that we have [LLMs] supporting development, the difficulty of the language doesn't matter. I'm seeing a huge opportunity behind strong runtime languages like Rust. [..] In [Developing Country] we hire 20-25 high school graduates, train them to be Rust programmers, then they enhance our workforce worldwide." -- Founder of a consulting company

We heard this from one organization. This is a claim that the combination of Rust's compiler and LLM tooling can dramatically shorten the path from beginner to working developer. Whether it generalizes depends on questions we can't answer from a single interview: how long these developers stay, what kind of code they can maintain independently, and whether this training/learning model works outside this company's particular structure. If it holds up, the pool of people who can become Rust developers is much larger than the usual hiring profile suggests.

Organizational considerations for Rust learners

We spoke with a number of folks on teams that are using Rust in larger organizations. Teams wanted to know that everyone would end up at roughly the same level of competence, which led a good number to invest in training courses to get there. Some leaders found that staff was able to ramp well enough by reading The Rust Programming Language, going through Rustlings, and then picking up lower risk and priority tickets to work on. Having a sense of community was also important within companies; it helps people know they are not alone when they are asked to work on Rust after, say, a reorganization happens.

"[..] the idea with the class as opposed to 'just read the Rust book on your own' was that this gives everyone kind of the same baseline going in." -- Principal Firmware Engineer (mobile robotics)

"So typically we're going to have people work through Rustlings, work through The Rust Programming Language. We have them then start to pick up lower risk tickets to work on." -- Principal Engineer at a large SaaS provider

"We've got an internal Slack channel for Rust learning where people can drop questions and others will come in and answer them. That helps build up understanding and community." -- Software Engineer at a large corporation

Some organizations found that while the person they'd hire would need to learn Rust, it was still preferable to the alternative of hiring someone for a critical piece of software written in another language.

"They needed to grow and maintain this C++ codebase. They had a C++ wizard, and they tried for about two years to find someone with the same level of expertise. They ended up hiring people that didn't know Rust and ramping them up, creating FFI bindings from the C++ side so they could work in Rust. And you can feel it: the borrow checker is teaching these people the right way to handle their systems." -- Principal Engineer at an Automotive OEM

The community and helping each other aspect seems to grow bonds as organizations mature.

"Our team is [all about] mentorship. I've mentored people coming up to speed on Rust, and people help each other hugely." -- Principal Software Engineer at a large SaaS company

Silent attrition

We identified some cases where people have approached Rust and bounced off of it, for one reason or another. In the below case, someone with a background in a language with fewer guardrails found themselves frustrated enough with Rust to walk away.

"All of that means that that embedded ecosystem is very frustrating to somebody who comes from C and is like, why can't I just get a pointer to this peripheral and then write into the registers. What are you doing to me? [..] My friend never got over that. He looked at it and said, I'm not going to deal with this and walked away." -– A second University Professor

There may be language features that for a particular domain are not seen as comfortable or usable yet, such as async Rust usage in a safety domain. We'd like to map which language features feel off-limits in which domains; async in safety-critical work probably isn't the only case.

"We're not fully sure how async [Rust] will work out in the long run in our domain. [..] People don't feel comfortable yet since C++14 doesn't provide such concepts. [..] It's the chicken-and-egg problem again: we probably need to gain some experience to see whether we can actually benefit from these new concepts in the automotive and safety domains." -- Team Lead at Automotive Supplier (ASIL D target)

We heard in at least one case, that while the language was challenging and there was a near bounce, the tooling helped keep them coming back and trying.

"Well, I think my early impressions of Rust - one is I find C++ so intimidating, and I think a big part of why I was able to succeed at [..] learning Rust is the tooling. I mean, all this makes sense [..] but it's like, for me, getting started with Rust, the language was challenging, but the tooling was incredibly easy." -- Founder of another startup built on Rust

While it might be considered more of a community concern, if there are interactions online and in spaces that point to learners having so-called "skill issues" this feeds into the narrative that Rust must be hard to learn. We may be unintentionally turning away Rust Project contributors and maintainers due to the vibes being put out when new learners show up in certain spaces.

"People are very helpful, but generally the attitude is: if your program is very complicated, it's mostly a skill issue. There's not that much empathy when people get stuck learning, and a lot of people are just pushed away by it. There's probably a huge number of people who silently stop wanting to write Rust, because at some point it gets complicated and the feedback they get is 'you just need to be a better programmer, obviously'." -- Software Engineer at a SaaS Provider

Feedback on near-bounces from survey

We found a few interesting perspectives collected in the Rust Vision doc survey which we administered with examples of bouncing and coming back:

"I started before 1.0, got stuck very soon when trying to translate patterns from C++ to Rust (due to borrow checking). I tried again after 1.0 and it stuck. [..]" -- Survey Respondent A

Survey Respondent A went on to share in a more detailed response about a perceived weakness in Rust learning materials related to lifetimes and the borrow checker are explained. There was an observation that it's fairly easy to run into more complex situations with lifetimes and the borrow checker. They felt that the current state of this sort of material and tutorials is fairly superficial and can leave learners stuck when they run into those more complex situations.

One respondent that bounced once and came back shared challenges around usage of async. In concert with Rust's memory-safety and the borrow checker, they found some of the nitty-gritty details of async were difficult to learn. While we're aware of the Rust Project's continuous efforts to improve Rust's async story, this is another data point of a user that faced challenges.

Another survey respondent shared how they had multiple times bounced in trying to learn Rust. They returned after a year or so and found Rustlings to be highly motivating. We note that having multiple pathways for folks to learn Rust opens up more possibilities for those that nearly bounced, just like this person.

Need more focused work on silent attritrion

The thing that stood out most to us was the lack of real, first-hand knowledge of having bounced when learning Rust. While this is an obvious effect of soliciting answers to our survey and opportunities to interview through Rust channels and our networks, this cohort is good future candidate where interviews could start.

Conclusions

Across these conversations, the experience of learning Rust depended heavily on context. Why someone was learning and what support they had mattered as much as the borrow checker. The same kinds of examples kept coming up: a training course that got a team to a shared baseline, a maintainer answering a student's first GitHub issue, and a colleague whose code showed that cloning was okay.

That context is largely something the community has a hand in. With that in mind, here is what we take away from what we heard, and what we still don't know.

What seems worth trying

Learning materials aimed at unlearning. Syntax barely came up when people described their struggles. People struggled with unlearning habits from previous languages, whether OOP structuring from C++ and Java or the instinct to grab a raw pointer to a peripheral. Most of our learning materials teach Rust from first principles, and that works. What we didn't come across is much written for, say, the engineer with ten years of Java who lands on a Rust team after a reorg: material that names the patterns they'll reach for that won't transfer, and shows what to do instead. The professor we spoke with did a version of this in the classroom, leaning on "traits are like interfaces in Java" and saving generics for later in the course, and the students did fine. Something similar could work outside the classroom too.

Put the "clone freely while you're learning" advice somewhere official. Every experienced developer we spoke with gave the same advice, but learners seem to mostly pick it up by accident, like the researcher who happened to see someone else cloning the struct they had been carefully threading lifetimes through. Saying it early in official materials would take some of the steepness out of the curve. The broader version belongs there too: idiomatic Rust doesn't have to mean optimal Rust, especially on a first project.

Diagnostics are already a primary learning resource: several people told us the compiler taught them lifetimes before any documentation did. Diagnostics reach learners right at the moment they're stuck. When writing new ones, it seems worth keeping the confused newcomer in mind alongside the expert, because for a lot of people this is where the learning happens.

Is "the book" actually out of date? Whether or not The Rust Programming Language or other materials are actually behind, a team evaluating Rust looked at its repository, saw unresolved issues and unmerged PRs, and moved on. As more companies evaluate adoption, more people will look at these materials with the same fresh eyes. Visible issue triage and some communication about what's current and what's planned would address the perception, separately from whatever content work may or may not be needed.

How stuck learners get treated is shaping who stays. We heard about students getting answers on GitHub from the maintainers who wrote the code, and we heard about learners being told their struggles were a skill issue. The first group came away with a lasting good impression of Rust. Some of the second group walked away entirely, and because they leave quietly, it's easy to underestimate how many of them there are. The welcoming side of the community came up unprompted as a reason people stayed, so we know it makes a difference when we get this right.

Every organization we spoke with described essentially the same ramp-up for bringing a team to Rust. Teams that brought groups of developers to Rust described roughly the same approach: get everyone to a shared baseline with a training course or with The Rust Programming Language and Rustlings, start people on lower-risk tickets, and give them somewhere internal to ask questions. Several organizations also found that hiring developers without Rust experience and ramping them up worked out better than continuing to search for rare expertise in another language. None of this is complicated, and teams weighing adoption don't need to invent a training program from scratch.

What we still don't know

The biggest gap is the people we didn't reach. Nearly everyone we spoke with stuck with Rust long enough to be reachable through Rust channels, so the stories of bouncing off came to us second-hand: a friend who walked away from embedded Rust, colleagues who quietly stopped after the responses they got. As we wrote in our first post, finding people who decided against Rust takes targeted outreach. If the proposed User Research team comes together, talking with learners who bounced would make a good early project, and learning is probably the area where that research would teach us the most.

We also don't know what to make of LLMs as a learning tool yet. They came up as a search engine, as an example generator, and in one organization's case as something that makes training high school graduates into working Rust developers possible. We saw a classroom where the C cohort leaned on LLMs in ways the Rust cohort couldn't, and we don't have an explanation for it. All of this comes from a handful of conversations, so we treat it as a set of leads to follow up on. Given how quickly the tools are changing, it seems better to study this deliberately than to wait and see what folklore develops.

The folks we spoke with showed that people do get there: with enough passes through the materials and enough code written, it eventually clicks. The opportunities above are mostly about making it work for the people who didn't pick Rust on purpose, and for the ones who would have stuck around if their early experience had gone a little differently.

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!

- In a world-changing breakthrough, Microsoft researchers have integrated OpenAI’s ChatGPT with robotics and other devices, enabling the use of natural language to control various devices, including drones, robot arms, home assistant robots and more, allowing humans to easily communicate their intentions to robots just like they would to another human, bridging the gap between man and machine. Tesla’s Optimus AI robot can now walk short distances by itself, carry an object of its arm's size, and use basic tools with the help of artificial intelligence. CACTI from Meta AI consists of four stages that handle data collection, data augmentation, visual representation learning, and imitation policy training separately to enable scalability in robot learning.

AI Marketplace: https://taimine.com/
Deep Learning AI Specialization: https://imp.i384100.net/GET-STARTED

AI News Timestamps:
0:00 New OpenAI Robotics Tech With ChatGPT
5:06 Tesla Optimus AI Robot News
6:15 Meta AI CACTI Robot Imitation Learning

#technology #tech #ai

The future of AI chips might not run on electricity. It might run on light.

In this video, we break down how Lightmatter is using photonic computing to attack one of the biggest problems in artificial intelligence: moving massive amounts of data between chips without wasting huge amounts of power. Lightmatter’s Envise processor uses light to perform AI calculations such as matrix multiplication, while its Passage platform uses optical interconnects to move data between AI chips at extreme speeds. The company has demonstrated real AI workloads including ResNet, BERT, and reinforcement learning models, while targeting the growing power and bandwidth crisis inside modern AI data centers. We also explore the Passage M1000 photonic superchip, Lightmatter’s collaboration with Qualcomm, its involvement in NVIDIA’s NVLink Fusion ecosystem, and why optical computing could become critical as AI clusters continue scaling. Could light-powered AI chips replace traditional electronic computing, or will photonics first transform how GPUs communicate within massive data centers?

#AIChips #Lightmatter #PhotonicComputing #Nvidia #ArtificialIntelligence #DataCenter #FutureTech

Google just revealed two brand-new AI chips—and they could change the future of artificial intelligence. Instead of building one GPU to handle everything, Google introduced its 8th-generation Tensor Processing Units: TPU 8t for AI training and TPU 8i for AI inference and reasoning. The company believes future AI infrastructure needs specialized hardware rather than one general-purpose accelerator. In this video, we break down Google's new TPU architecture, including 121 exaflops of compute, superpods with up to 9,600 chips, 2 petabytes of shared memory, Virgo networking, TPUDirect, Boardfly topology, Axion CPUs, and next-generation liquid cooling. We also explain why Google optimized TPU 8i for reasoning models, AI agents, and massive inference workloads with 288GB of HBM and dramatically improved memory performance. Could Google's specialized TPU strategy become the future of AI computing? And is this the first real architectural challenge to NVIDIA's GPU dominance?

#Google #TPU #NVIDIA #AIChips #ArtificialIntelligence #GoogleCloud #Semiconductors

NVIDIA just unveiled Rubin, but this isn't just another GPU. It's a complete AI factory built to power the next generation of artificial intelligence. In this video, we break down NVIDIA's Vera Rubin platform, a fully co-designed AI system that combines the Vera CPU, Rubin GPU, NVLink 6, BlueField-4 DPU, ConnectX-9 SuperNIC, and Spectrum networking into one massive AI infrastructure platform. You'll learn why Rubin delivers up to 50 petaflops of AI inference, 288GB of HBM4 memory per GPU, 3.6 exaflops of compute per NVL72 rack, over 260TB/s of NVLink bandwidth, and why NVIDIA claims up to 10x lower AI inference costs than Blackwell. We also explore why AI companies like Microsoft, OpenAI, Anthropic, Meta, xAI, Google, AWS, Oracle, and CoreWeave are already adopting Rubin, and why NVIDIA believes the future of AI isn't a faster GPU; it's an entire factory built to manufacture intelligence. Is Rubin the biggest leap in AI hardware yet, or the beginning of a completely new era of AI infrastructure?

#NVIDIA #Rubin #AIFactory #AIChips #ArtificialIntelligence #HBM4 #DataCenter

NVIDIA just revealed the most powerful AI workstation ever built—and it puts data center hardware on your desk. Powered by the new GB300 Grace Blackwell Ultra Superchip, the NVIDIA DGX Station combines a 72-core Grace CPU, a Blackwell Ultra GPU with 20,480 CUDA cores, 748GB of unified coherent memory, and up to 20 petaflops of AI compute. It's designed to run massive AI models locally, eliminating many of the memory limitations that force developers to rely on expensive cloud GPUs. In this video, we break down the DGX Station architecture, unified memory, NVLink C2C, HBM3e, local AI inference, trillion-parameter model claims, real-world pricing, and why NVIDIA believes desktop AI workstations are the future of artificial intelligence development. We also compare the DGX Station with DGX Spark, Apple’s Mac Studio, cloud GPU infrastructure, and explain why local AI could become the next major shift in computing.

Is NVIDIA reinventing the personal computer for the AI era?

#NVIDIA #DGXStation #Blackwell #AIWorkstation #ArtificialIntelligence #LocalAI #CUDA

The global AI race just entered a dangerous new phase. China is reportedly considering restricting access to its most advanced AI models; DeepSeek is developing its own AI chip; ByteDance and Alibaba are shutting down AI companion services under new regulations, and the Claude Code controversy has escalated into a geopolitical dispute. At the same time, both China and the United States are tightening control over frontier AI technologies, turning artificial intelligence into a strategic national asset. In this video, we break down the biggest AI developments of the week, including China's AI export discussions, DeepSeek's custom silicon ambitions, Huawei-powered AI innovation, Anthropic's Claude Code controversy, MoWorld's real-time AI world model, and why many experts believe we're witnessing the beginning of an AI Cold War. Is the era of open AI coming to an end?

#AI #China #ArtificialIntelligence #DeepSeek #OpenAI #Anthropic #NVIDIA #Huawei

As a small boy growing up in Amritsar, India, in the 1960s, Avtar Singh used to hear whispers. Your mum isn’t really your mum, the rumours would say. Your mum lives abroad. Avtar didn’t really pay any attention to it. “I thought, the neighbours are just making up stories.” But the stories were true: the woman raising Avtar was actually his grandmother. When he was too young to remember it, his father had emigrated to Canada to work as a teacher, leaving Avtar in the care of his paternal grandparents. His grandfather, a police officer, was a fearsome authoritarian, but his grandmother brought him up with tenderness, as if he were her own son. She called him by his nickname, Titu. He felt loved.

When Avtar was around eight years old, he was told about his family in Canada: his dad, his mum and a little brother were waiting for him there, and he would soon travel to join them. Just after his ninth birthday, Avtar found himself sitting on a plane, dressed in a three‑piece suit and tie. He was travelling alone. He had never flown before. He didn’t speak a word of English. He landed in Halifax, Nova Scotia, on Christmas Eve, 1968; his family met him at the airport and took him home. “I didn’t know what to think,” Avtar, now 66, tells me in a video call from his home in Abbotsford, British Columbia. “Before I knew it, it was the first week of January, and I was put in school.”

Every weekday, Matt, a software engineer, looks forward to his four-hour train commute to Pawling, New York. It’s time he uses to work on his own project: a browser-based video game for which he writes every line of code himself.

“I am actively trying to keep my axe sharp,” said Matt, who did not want to use his actual name, to protect his employment. In the last six months, Matt’s job has increasingly shifted away from coding, problem solving and software architecture towards reviewing code generated by artificial intelligence. Convinced that the shift will weaken his skills, he’s doing what he can to keep them intact. “I am trying not to leverage AI where I can.”

Dieser Beitrag RefluXFS: Neue XFS-Kernel-Schwachstelle ermöglicht Root-Rechte auf betroffenen RHEL-Installationen erschien als erstes auf IT BOLTWISE x Artificial Intelligence.

Read more of this story at Slashdot.

Dieser Beitrag Check Point schließt SmartConsole-Lücke (CVE-2026-16232) nach aktiver Ausnutzung erschien als erstes auf IT BOLTWISE x Artificial Intelligence.

Read more →

The post Claude Mythos Just Caught the Attention of Canada’s Banking Regulator appeared first on IT Security News.

Read more of this story at Slashdot.

The post Zuckerberg Wants to Change How You Think About AI appeared first on TechRepublic.

Dieser Beitrag Warum KI-Projekte oft scheitern: Incentives statt Modelle im Fokus erschien als erstes auf IT BOLTWISE x Artificial Intelligence.

Read more →

The post OpenAI Models Breach Hugging Face During Cyber Test appeared first on IT Security News.

Read more →

The post ChatGPT AgentForger Flaw Could Deploy Rogue Workspace Agents via a Phishing Link appeared first on IT Security News.

Read more →

The post The AI Trust Paradox: Businesses Are Racing Ahead, but Consumers Are Hesitating appeared first on IT Security News.

Zimbra Phishing Campaign Uses CVE-2025-66376

LAUNDRY BEAR Targets Email and Sensitive Data

Ulej and Flowerbed Support Email Data Exfiltration

Organizations Urged to Patch Vulnerable Zimbra Systems

Dieser Beitrag HermeticReader: Schwachstelle in Adobe-Acrobat-Erweiterung liest WhatsApp-Web-Daten mit UXSS erschien als erstes auf IT BOLTWISE x Artificial Intelligence.

Dieser Beitrag GitHub senkt öffentliche Bug-Bounty-Zahlungen und verlagert Top-Prämien in eine VIP-Stufe erschien als erstes auf IT BOLTWISE x Artificial Intelligence.

Dieser Beitrag CVE-2026-8933: Ubuntu snap-confine ermöglicht lokalen Root-Zugriff auf Standard-Installationen erschien als erstes auf IT BOLTWISE x Artificial Intelligence.

The announcement, made by Microsoft AI's Superintelligence team, lands roughly a year after the company committed to building purpose-built models internally, and it arrives with an unusual level of specificity about where those models now run: Bing, PowerPoint, OneDrive, Dynamics 365, Excel, GitHub Copilot, and Azure. The message to enterprise buyers — and, implicitly, to OpenAI — is that Microsoft's homegrown models are no longer research projects. They are production infrastructure serving millions of users.

"Each of these enhancements is a step toward the same goal: Microsoft products, powered by Microsoft models," the company wrote in its announcement blog.

How MAI-Image-2.5-Pro and MAI-Voice-2-Flash stake out opposite ends of the AI cost curve

The two new releases occupy opposite ends of what Microsoft calls the quality-speed-cost curve, and the positioning is deliberate. MAI-Image-2.5-Pro targets the premium tier: hero imagery, detailed editing, and precise in-image text rendering — the last of which has long been a notorious weak spot for image generation models. Microsoft priced the model at $5 per million text input tokens, $8 per million image input tokens, and $106 per million image output tokens. The base MAI-Image-2.5 model recently launched at No. 2 for image editing on Arena, the community leaderboard that has become a de facto scoreboard for generative media.

The creative industry appears to be taking notice. Rob Reilly, global chief creative officer at advertising giant WPP, called the Pro model "a strong leap forward for GenMedia tools" in a statement included in Microsoft's announcement, adding that "Microsoft has firmly established itself among the leaders in generative AI."

MAI-Voice-2-Flash goes the other direction. First previewed at Microsoft's Build conference, Flash runs twice as fast as MAI-Voice-2 and costs 32% less, priced at $15 per million characters. It is designed for the unglamorous but enormous market of high-volume voice — call centers, voice agents, and real-time speech applications where latency and cost-per-call matter more than marginal gains in expressiveness. Together, the two models reflect a strategy of building families of models rather than a single flagship, because, as the company put it, a creative studio chasing maximum fidelity has very different needs from a customer service operation handling millions of calls a day.

Microsoft's production metrics show in-house models cutting GPU costs by up to 89%

The model launches are arguably less newsworthy than the deployment metrics Microsoft attached to them — numbers that read like a systematic case for swapping out third-party frontier models across its product portfolio. 

Bing Image Creator now runs entirely on MAI-Image-2.5, end to end, marking the first time the consumer image tool is fully in-house. In PowerPoint, Microsoft says MAI-Image-2.5 reduces GPU costs by up to 84% compared with GPT-Image-2, OpenAI's image model. In OneDrive, where MAI-Image-2.5 is now the default for key image-editing scenarios, the company reports a 26% increase in save rates, roughly 25% lower P95 latency, and 2.5 times greater efficiency under medium-utilization production workloads.

On the voice side, MAI-Voice-2-Flash now powers Dynamics 365 Contact Center — the platform used by customers including T-Mobile and EasyJet — where Microsoft claims GPU cost reductions of up to 89%. The model is also integrated into Azure Voice Live for developers building speech-to-speech agents.

Perhaps the most consequential deployment sits in healthcare. Microsoft's Dragon Copilot, used by 170,000 medical providers and responsible for processing 28 million patient encounters last quarter, now runs on MAI-Transcribe-1.5 for its multilingual workflow across 58 languages. Microsoft says internal evaluations show a 50% relative reduction in both transcription and language-identification error rates across most languages — a meaningful claim in a domain where transcription errors can propagate directly into clinical notes.

Inside the 'hill-climbing' strategy that lets small models beat GPT-5.6 in Excel

In a companion post published the same day, Microsoft detailed the methodology behind these results — what it calls its "hill-climbing machine," an integrated flywheel of data, models, and the product "harness" that surrounds them.

The clearest example is MAI-Code-1-Flash, the lightweight coding model launched in GitHub Copilot in June. Microsoft says the model achieves an approximately 10% higher code accept rate than GPT-5.4 Mini and Claude Haiku 4.5 in VS Code, while using 10% fewer median tokens. Developer retention tells a similar story: users were 6% more likely to return across multiple days than with GPT-5.4 Mini, and 11% more likely than with Claude Haiku 4.5.

Then Microsoft did something more interesting. It took the MAI-Code-1-Flash checkpoint and further trained it inside an Excel reinforcement learning environment, teaching a coding model the tools and workflows of spreadsheet knowledge work. The result, according to production user feedback, is a model on par with GPT-5.6 for the most common Excel tasks — while being small enough to run on Nvidia's older H100 and even A100 GPUs rather than requiring the latest-generation accelerators.

That hardware detail deserves emphasis. Every major AI company is fighting for allocation of cutting-edge chips, and a model that delivers frontier-adjacent quality on two-generation-old silicon fundamentally changes the deployment economics. It also frees the newest hardware — including Microsoft's now-operational GB200 cluster — for training rather than serving.

Satya Nadella's 'frontier diffusion' manifesto redraws the OpenAI relationship

Microsoft CEO Satya Nadella framed the announcements in a lengthy post on X titled "Frontier Diffusion & Control," which functions as something close to a strategic manifesto. "We can now take saturated frontier capabilities and deliver them at scale and at lower cost through models optimized for high-usage products, while continuing to use frontier models for frontier needs," Nadella wrote, adding that Microsoft is "beginning to route traffic across our first-party surfaces to MAI whenever our models match or outperform frontier alternatives."

Translated from executive prose: capabilities that were state-of-the-art a year ago are now table stakes, and Microsoft believes it can replicate them cheaply for the specific, repetitive tasks that dominate real product usage. Why pay frontier prices for a frontier model when a user just wants to reformat a spreadsheet column?

Nadella was careful to note that "frontier models from OpenAI and Anthropic are part of the orchestration system alongside MAI" — but he also articulated a pointed principle of model independence, arguing that a company's evaluations "should continue to hill climb even when any given model has been removed." 

“Keeping the harness, memory, context, and skills outside the model, he argued, is what gives Microsoft control. The subtext is hard to miss. Reuters reported in April that Microsoft’s exclusive license to OpenAI’s technology had been revised into a non-exclusive arrangement, and The Information reported last September that Microsoft had begun incorporating Anthropic models into some products. Wednesday’s announcement completes the triangle: Microsoft as orchestrator, with its partners’ frontier models as interchangeable components and its own models absorbing an ever-larger share of routine traffic.”

Developers cheer cheaper task-specific models while skeptics question Microsoft's track record

The response online captured both the appeal and the skepticism surrounding the strategy. "I love when people use small models for niche tasks," wrote one X user, @mavihsk, responding to Nadella's post. "Why do I have to use the all-knowing model just to change my field in Excel?" Another user, @nabu_lines, distilled the pitch neatly: "cost and performance both improve when you stop overusing the biggest model."

Others were less charitable about Microsoft's execution track record. "Microsoft is the worst when it comes to listening to user feedback," wrote designer @designedbyabin, arguing the company "will lose the AI race because they repeatedly failed to understand user needs." And one user, @tokenoverflow, offered a drier critique of the model-independence pitch: "i want it keep hill climbing after removing microsoft."

The skeptics raise a fair point. Microsoft's self-reported metrics — accept rates, save rates, GPU savings — come from its own internal evaluations, not independent benchmarks, and the company chooses which comparisons to publish.

But the strategy's logic does not depend on any single number. Nadella's framing that software now has "real marginal cost for the first time" explains why Microsoft is obsessive about tokens, GPUs, and serving costs: when AI features run on every keystroke across a billion-user product portfolio, an 84% GPU cost reduction is not an optimization. It is the difference between a viable business and a money pit.

Why Microsoft is turning its internal AI playbook into an Azure product

The final piece of the strategy is that Microsoft is selling the playbook, not just the models. Nadella explicitly positioned the hill-climbing approach as "a template for every other AI native, SaaS, or Enterprise company," and Microsoft is packaging the toolchain through Foundry and what it calls Frontier Tuning — letting enterprises train specialized models against their own proprietary evaluations and reinforcement learning environments. That turns Microsoft's internal cost-cutting exercise into an Azure product, and it gives enterprise customers a reason to run their AI workloads on Microsoft's cloud even if the models themselves come from elsewhere.

The company's emphasis on models trained "on clean, traceable, enterprise-grade data, without distillation from third-party models" serves the same commercial end. In an industry facing mounting scrutiny over training data provenance, Microsoft is betting that enterprise buyers — and courts — will care where model capabilities come from. Microsoft says it is now extending the hill-climbing approach to Copilot Chat, Outlook, and PowerPoint, and both new models are available in public preview through Microsoft Foundry and the MAI Playground. "None of this is an endpoint," the company wrote. "We're just getting started."

Seven years ago, Microsoft bet more than $13 billion that OpenAI would build the future of AI. Wednesday's announcement suggests the company has since learned a cheaper lesson: the future of AI may belong to whoever builds the frontier, but the profits belong to whoever makes it ordinary.

Dieser Beitrag KI-Adoption im Unternehmen: Security gewinnt Tempo durch Governance erschien als erstes auf IT BOLTWISE x Artificial Intelligence.

Dieser Beitrag Windmill-Kritik: CVE-2026-29059 liest Serverdateien ohne Authentifizierung erschien als erstes auf IT BOLTWISE x Artificial Intelligence.

Dieser Beitrag KI-Einführung gelingt schneller, wenn Security Governance als Enablement denkt erschien als erstes auf IT BOLTWISE x Artificial Intelligence.

Dieser Beitrag Warum moderne SOCs Multi-Layer-Detektionen und NDR brauchen erschien als erstes auf IT BOLTWISE x Artificial Intelligence.

Read more →

The post Don’t swing at everything appeared first on IT Security News.

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.

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.

This article Copilot hacks for work and productivity first appeared on TheWindowsClub.com.

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.

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:

1. sendStartPing,
2. gather_email,
3. gather_environment,
4. gather_2fa_codes,
5. gather_app_password,
6. gather_device_status,
7. gather_oauth_consumers,
8. gather_autocomplete_password,
9. enable_mail_protocols,
10. gather_gal,
11. sendArchives, and
12. sendFinishPing. 

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.

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.

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.

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.

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.  

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.  

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.

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.

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.

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].

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
² Forsvarets Efterretningstjeneste
³ Välisluureamet
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
¹¹ Agencja Wywiadu
¹² Służba Kontrwywiadu Wojskowego
¹³ Centro Nacional de Inteligencia
¹⁴ Nationellt 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

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.