Read more of this story at Slashdot.

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.

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.

Read more of this story at Slashdot.

Note: This joint Cybersecurity Advisory is being published as an addition to the Cybersecurity and Infrastructure Security Agency (CISA) May 6, 2025, joint fact sheet Primary Mitigations to Reduce Cyber Threats to Operational Technology and European Cybercrime Centre’s (EC3) Operation Eastwood, in which CISA, Federal Bureau of Investigation (FBI), Department of Energy (DOE), Environmental Protection Agency (EPA), and EC3 shared information about cyber incidents affecting the operational technology (OT) and industrial control systems (ICS) of critical infrastructure entities in the United States and globally.

FBI, CISA, National Security Agency (NSA), and the following partners—hereafter referred to as “the authoring organizations”—are releasing this joint advisory on the targeting of critical infrastructure by pro-Russia hacktivists:

• U.S. Department of Energy (DOE)
• U.S. Environmental Protection Agency (EPA)
• U.S. Department of Defense Cyber Crime Center (DC3)
• Europol European Cybercrime Centre (EC3)
• EUROJUST – European Union Agency for Criminal Justice Cooperation
• Australian Signals Directorate’s Australian Cyber Security Centre (ASD’s ACSC)
• Canadian Centre for Cyber Security (Cyber Centre)
• Canadian Security Intelligence Service (CSIS)
• Czech Republic Military Intelligence (VZ)
• Czech Republic National Cyber and Information Security Agency (NÚKIB)
• Czech Republic National Centre Against Terrorism, Extremism, and Cyber Crime (NCTEKK)
• French National Cybercrime Unit – Gendarmerie Nationale (UNC)
• French National Jurisdiction for the Fight Against Organized Crime (JUNALCO)
• German Federal Office for Information Security (BSI)
• Italian State Police (PS)
• Latvian State Police (VP)
• Lithuanian Criminal Police Bureau (LKPB)
• New Zealand National Cyber Security Centre (NCSC-NZ)
• Romanian National Police (PR)
• Spanish Civil Guard (GC)
• Spanish National Police (CNP)
• Swedish Polisen (SC3)
• United Kingdom National Cyber Security Centre (NCSC-UK)

The authoring organizations assess pro-Russia hacktivist groups are conducting less sophisticated, lower-impact attacks against critical infrastructure entities, compared to advanced persistent threat (APT) groups. These attacks use minimally secured, internet-facing virtual network computing (VNC) connections to infiltrate (or gain access to) OT control devices within critical infrastructure systems. Pro-Russia hacktivist groups—Cyber Army of Russia Reborn (CARR), Z-Pentest, NoName057(16), Sector16, and affiliated groups—are capitalizing on the widespread prevalence of accessible VNC devices to execute attacks against critical infrastructure entities, resulting in varying degrees of impact, including physical damage. Targeted sectors include Water and Wastewater Systems, Food and Agriculture, and Energy.

The authoring organizations encourage critical infrastructure organizations to implement the recommendations in the Mitigations section of this advisory to reduce the likelihood and impact of pro-Russia hacktivist-related incidents. For additional information on Russian state-sponsored malicious cyber activity, see CISA’s Russia Threat Overview and Advisories webpage.

Download the PDF version of this report:

Background and Development of Pro-Russia Hacktivist Groups

Over the past several years, the authoring organizations have observed pro-Russia hacktivist groups conducting cyber operations against numerous organizations and critical infrastructure sectors worldwide. The escalation of the Russia-Ukraine conflict in 2022 significantly increased the number of these pro-Russia groups. Consisting of individuals who support Russia’s agenda but lack direct governmental ties, most of these groups target Ukrainian and allied infrastructure. However, among the increasing number of groups, some appear to have associations with the Russian state through direct or indirect support.

Cyber Army of Russia Reborn

The authoring organizations assess that the Russian General Staff Main Intelligence Directorate (GRU) Main Center for Special Technologies (GTsST) military unit 74455—tracked in the cybersecurity community under several names (see Appendix B: Additional Designators Used for Cited Groups)—is likely responsible for supporting the creation of CARR —also known as “The People’s Cyber Army of Russia”—in late February or early March of 2022. Actors suspected to be from GRU unit 74455 likely funded the tools CARR threat actors used to conduct distributed denial-of-service (DDoS) attacks through at least September 2024.

In April 2022, the group began using a new Telegram channel featuring the name “CyberArmyofRussia_Reborn” to organize and plan group actions. The channel creators recruited actors to use CARR as an unattributable platform for conducting cyber activities beneath the level of an APT, aimed at deterring anti-Russia rhetoric. CARR threat actors presented themselves as a group of pro-Russia hacktivists supporting Russia’s stance on the Ukrainian conflict, and they soon began claiming responsibility for DDoS attacks against the U.S. and Europe for supporting Ukraine.

CARR documented these actions through embellished images and videos shared on their social media channels, promoting Russian ideology, disseminating talking points, and publicizing leaked information from hacks attributed to Russian state threat actors.

In late 2023, CARR expanded their operations to include attacks on industrial control systems (ICS), claiming an intrusion against a European wastewater treatment facility in October 2023. In November 2023, CARR targeted human-machine interface (HMI) devices, claiming intrusions at two U.S. dairy farms.

The authoring organizations assess that by late September 2024, CARR channel administrators became dissatisfied with the level of support and funding provided by the GRU. This dissatisfaction led CARR administrators and an administrator from another hacktivist group, NoName057(16), to create the Z-Pentest group, employing the same tactics, techniques, and procedures (TTPs) as CARR but separate from GRU involvement.

NoName057(16)

The authoring organizations assess that the Center for the Study and Network Monitoring of the Youth Environment (CISM), established on behalf of the Kremlin, created NoName057(16) as a covert project within the organization. Senior executives and employees within CISM developed and customized the NoName057(16) proprietary DDoS tool DDoSia, paid for the group’s network infrastructure, served as administrators on NoName057(16) Telegram channels, and selected DDoS targets.

Active since March 2022, NoName057(16) has conducted frequent DDoS attacks against government and private sector entities in North Atlantic Treaty Organization (NATO) member states and other European countries perceived as hostile to Russian geopolitical interests. The group operates primarily through Telegram channels and used GitHub, alongside various websites and repositories, to host DDoSia and share materials and TTPs with their followers. 

In 2024, NoName057(16) began collaborating closely with other pro-Russia hacktivist groups, operating a joint chat with CARR by mid-2024. In July 2024, NoName057(16) jointly claimed responsibility with CARR for an alleged intrusion against OT assets in the U.S. The high degree of cooperation with CARR likely contributed to the formation of Z-Pentest, which is composed of actors and administrators from both teams, in September 2024.

Z-Pentest

Established in September 2024, Z-Pentest is composed of members from CARR and NoName057(16). The group specializes in OT intrusion operations targeting globally dispersed critical infrastructure entities. Additionally, the group uses “hack and leak” operations and defacement attacks to draw attention to their pro-Russia messaging. Unlike other pro-Russia hacktivist groups, Z-Pentest largely avoids DDoS activities, claiming OT intrusions as attempts to garner more attention from the media.

Shortly after Z-Pentest’s inception, the group announced alliances with CARR and NoName057(16), possibly to leverage the other groups’ subscribers to grow the new channel. In March 2025, Z-Pentest posted evidence claiming OT device intrusions to their channel using a NoName057(16) cyberattack campaign hashtag. Similarly, in April 2025, Z-Pentest shared a video purporting defacement of an HMI by changing system names to NoName057(16) and CARR references. Z-Pentest continues to create new alliances with other groups, like Sector16, to continue growing their subscriber base and incidentally propagate TTPs with new partners.

Sector16

Formed in January 2025, Sector16 is a novice pro-Russia hacktivist group that emerged through collaboration with Z-Pentest. Sector16 actively maintains an online presence, including a public Telegram channel where they share videos, statements, and claims of compromising U.S. energy infrastructure. These communications often align with pro-Russia narratives and reflect their self-proclaimed support for Russian geopolitical objectives.

Members of Sector16 may have received indirect support from the Russian government in exchange for conducting specific cyber operations that further Russian strategic goals. This aligns with broader Russian cyber strategies that involve leveraging non-state threat actors for certain cyber activities, adding a layer of deniability.

Technical Details

Note: This advisory uses the MITRE ATT&CK^® Matrix for Enterprise framework, version 18. See the MITRE ATT&CK Tactics and Techniques section of this advisory for a table of the threat actors’ activity mapped to MITRE ATT&CK tactics and techniques.

TTP Overview

Pro-Russia hacktivist groups employ easily disseminated and replicated TTPs across various entities, increasing the likelihood of widespread adoption and escalating the frequency of intrusions. These groups have limited capabilities, frequently misunderstanding the processes they aim to disrupt. Their apparent low level of technical knowledge results in haphazard attacks where actors intend to cause physical damage but cannot accurately anticipate actual impact. Despite these limitations, the authoring organizations have observed these groups willfully cause actual harm to vulnerable critical infrastructure.

Pro-Russia hacktivist groups use the TTPs in this Cybersecurity Advisory to target virtual network computing (VNC)-connected HMI devices. These groups are primarily seeking notoriety with their actions. While they have caused damage in some instances, they regularly make false or exaggerated claims about their attacks on critical infrastructure to garner more attention. They frequently misrepresent their capabilities and the impacts of their actions, portraying minor incursions as significant breaches, but such incursions can still lead to lost time and resources for operators remediating systems.

Additionally, pro-Russia hacktivists use an opportunistic targeting methodology. They leverage superficial criteria, such as victim availability and existing vulnerabilities, rather than focusing on strategically significant entities. Their lack of strategic focus can lead to a broad array of targets, ranging from water treatment facilities to oil well systems. Pro-Russia hacktivists have demonstrated a pattern of frequently taking advantage of the widespread availability of vulnerable VNC connections. While system owners typically use VNC connections for legitimate remote system access functions, threat actors can maliciously use these connections to broadly target numerous platforms and services. Consequently, these groups can indiscriminately compromise critical infrastructure entities, including those in the Water and Wastewater, Food and Agriculture, and Energy Sectors.

Pro-Russia hacktivist groups have successfully targeted supervisory control and data acquisition (SCADA) networks using basic methods, and in some cases, performed simultaneous DDoS attacks against targeted networks to facilitate SCADA intrusions. As recently as April 2025, threat actors used the following unsophisticated TTPs to access networks and conduct SCADA intrusions:

• Scan for vulnerable devices on the internet [T0883] with open VNC ports [T1595.002].
• Initiate temporary virtual private server (VPS) [T1583.003] to execute password brute force software.
• Use VNC software to access hosts [T1021.005].
• Confirm connection to the vulnerable device [T0886].
• Brute force the password, if required [T1110.003].
• Gain access to HMI devices [T0883], typically with default [T0812], weak, or no passwords [T0859].
• Log the confirmed vulnerable device IP address, port, and password.
• Using the HMI graphical interface [T0823], capture screen recordings or intermittent screenshots while conducting the following actions, intending to affect productivity and cause additional costs [T0828]:
  • Modify usernames/passwords [T0892];
  • Modify parameters [T0836];
  • Modify device name [T0892];
  • Modify instrument settings [T0831];
  • Disable alarms [T0878];
  • Create loss of view (a technique that mandates local hands-on operator intervention) [T0829]; and/or
  • Device restart or shutdown [T0816].
• Disconnect from the device, ending the VNC connection.
• Research the compromised device company after the intrusion [T1591].

Propagation

To reach a wider audience, pro-Russia hacktivist groups work together, amplify each other’s posts, create additional groups to amplify their own posts, and likely share TTPs. For example, Z-Pentest jointly claimed intrusion of a U.S. system with Sector16. Sector16 later began posting additional intrusions for which the group claimed sole responsibility. It is likely that these and similar groups will continue to iterate and share these methods to disrupt critical infrastructure organizations.

Reconnaissance and Initial Access

The threat actors’ intrusion methodology is relatively unsophisticated, inexpensive to execute, and easy to replicate. These pro-Russia hacktivist groups abuse popular internet-scraping tools, such as Nmap or OPENVAS, to search for visible VNC services and use brute force password spraying tools to access devices via known default or otherwise weak credentials. Threat actors typically search for these services on the default port 5900 or other nearby ports (5901-5910). Their goal is to gain remote access to HMI devices connected to live control networks.

Once threat actors obtain access, they manipulate available settings from the graphical user interface (GUI) on the HMI devices, such as arbitrary physical parameter and setpoint changes, or conduct defacement activities. Because pro-Russia hacktivist groups seem to lack sector-specific expertise or cyber-physical engineering knowledge, they currently cannot reliably estimate the true impact of their actions. Regardless of outcome, pro-Russia hacktivist groups often post images and screen recordings to their social media platforms, boasting the compromises and exaggerating impacts to garner attention from their peers and the media.

Impact

While pro-Russia hacktivist groups currently demonstrate limited ability to consistently cause significant impact, there is a risk that their continued attacks will result in further harm or grievous physical consequences. Attacks have not yet caused injury; however, the attacks against occupied factories and community facilities demonstrate a lack of consideration for human safety.

Victim organizations reported that the most common operational impact caused by these threat actors is a temporary loss of view, necessitating manual intervention to manage processes. However, any modifications to programmatic and systematic procedures can result in damage or disruption, including substantial labor costs from hiring a programmable logic controller programmer to restore operations, costs associated with operational downtime, and potential costs for network remediation.

MITRE ATT&CK Tactics and Techniques

See Table 1 to Table 10 for all referenced threat actor tactics and techniques in this advisory. For assistance with mapping malicious cyber activity to the MITRE ATT&CK framework, see CISA and MITRE ATT&CK’s Best Practices for MITRE ATT&CK Mapping and CISA’s Decider Tool.

Incident Response

If organizations find exposed systems with weak or default passwords, they should assume threat actors compromised the system and begin the following incident response protocols:

1. Determine which hosts were compromised and isolate them by quarantining or taking them offline.
2. Initiate threat hunting activities to scope the intrusion. Collect and review artifacts, such as running processes/services, unusual authentications, and recent network connections.
3. Reimage compromised hosts.
4. Provision new account credentials.
5. Report the compromise to CISA, FBI, and/or NSA. See the Contact Information section of this advisory.
6. Harden the network to prevent additional malicious activity. See the Mitigations section of this advisory for guidance.

Mitigations

OT Asset Owners and Operators

The authoring organizations recommend organizations implement the mitigations below to improve your organization’s cybersecurity posture based on the threat actors’ activity. These mitigations align with the Cross-Sector Cybersecurity Performance Goals (CPGs) developed by CISA and the National Institute of Standards and Technology (NIST). The CPGs provide a minimum set of practices and protections that CISA and NIST recommend all organizations implement. CISA and NIST based the CPGs on existing cybersecurity frameworks and guidance to protect against the most common and impactful threats, tactics, techniques, and procedures. Visit CISA’s CPGs webpage for more information on the CPGs, including additional recommended baseline protections.

• Reduce exposure of OT assets to the public-facing internet. When connected to the internet, OT devices are easy targets for malicious cyber threat actors. Many devices can be found by searching for open ports on public IP ranges with search engine tools to target victims with OT components [CPG 3.S].
  • Asset owners should use attack surface management services and web-based search platforms to scan the internet. This mitigation can help identify if there are VNC systems exposed within the IP ranges they own, especially for connections set up by third parties.
    Note: For more information on attack surface management, see CISA’s Internet Exposure Reduction Guidance, CISA’s Cyber Hygiene Services for U.S. critical infrastructure, and NSA’s Attack Surface Management for the U.S. Defense Industrial Base.
  • Implement network segmentation between IT and OT networks. Segmenting critical systems and introducing a demilitarized zone (DMZ) for passing control data to enterprise logistics reduces the potential impact of cyber threats and the risk of disruptions to essential OT operations [CPG 3.I].
  • Consider implementing a firewall and/or virtual private network if exposure to the internet is necessary for controlling access to devices.
    • Consider disabling public exposure by default and implementing time-limited remote access to reduce the amount of time systems are exposed.
    • Restrict and monitor both inbound and outbound traffic at OT perimeter firewalls. Configure OT perimeter firewalls to enforce a default-deny policy for all traffic. Asset owners should explicitly permit authorized destinations and protocols based on operational requirements.
    • Implement strict egress filtering to prevent unauthorized data exfiltration or command-and-control callbacks.
    • Regularly audit firewall rulesets and monitor outbound traffic patterns for anomalies indicative of threat actor activity, such as beaconing or unexpected protocol usage.
• Adopt mature asset management processes, including mapping data flows and access points. Generating a complete picture of both OT and IT assets provides visibility to operators and management, allowing organizations to monitor and assess deviations for criticality [CPG 2.A].
  • Keep remote access services updated with the latest version available and ensure all systems and software are up to date with patches and necessary security updates.
    • Keep VNC systems updated with the latest version available.
  • Refer to the joint Foundations for OT Cybersecurity: Asset Inventory Guidance for Owners and Operators to help with reducing cybersecurity risk by identifying which assets within their environment should be secured and protected.
• Ensure OT assets use robust authentication procedures.
  • Many devices lack robust authentication and authorization. Devices with weak authentication are vulnerable targets to threat actors using credential theft techniques.
  • Implement MFA where possible. Where MFA is not feasible, use strong, unique passwords. Apply password standards for operator-accessible services on underlying OT assets, as well as network devices protecting those services. This is especially important for services that require internet accessibility [CPG 3.A] [CPG 3.B] [CPG 3.C] [CPG 3.F].
  • Establish an allowlist that permits only authorized device IP addresses and/or media access control addresses. The allowlist can be refined to operator working hours to further obstruct malicious threat actor activity; organizations are encouraged to establish monitoring and alerting for access attempts not meeting these criteria [CPG 3.E].
  • Disable any unused authentication methods, logic, or features, such as default authentication keys and default passwords. Block all unused high ephemeral ports and monitor for attempted connections using standard protocols on non-standard ports [CPG 3.R].
  • Authenticate all access to field controllers before authorizing access to, or modification of, a device’s state, logic, program, or filesystems.
• Enable control system security features that can separate and audit view and control functions. Limiting remotely accessible or default user accounts to “view-only” removes the potential for impact without exploiting a vulnerability [CPG 3.G].
• Implement and practice business recovery/disaster recovery plans. Plans should also take into consideration redundancy, fail-safe mechanisms, islanding capabilities, backup restoration, and manual operation.
  • Include scenarios that necessitate switching to manual operations. Maintaining the capability of an organization to revert to manual controls to quickly restore operations is vital in the immediate aftermath of a cyber incident [CPG 6.A].
  • Create backups of the engineering logic, configurations, and firmware of HMIs to enable fast recovery. Organizations should routinely test backups and standby systems to ensure safe manual operations in the event of an incident [CPG 3.O].
• Collect and monitor the traffic of OT assets and networking devices. This includes unusual logins or unexpected protocols communicating over the internet, and functions of ICS management protocols that change an asset’s operating mode or modify programs.
• Review configurations for setpoint ranges or tag values to stay within safe ranges and establish alerting for deviations.
• Take a proactive approach in the procurement process by following the guidance outlined in the joint guide Secure by Demand: Priority Considerations for Operational Technology Owners and Operators when Selecting Digital Products.

OT Device Manufacturers

Although critical infrastructure organizations can take steps to mitigate risks, it is ultimately the responsibility of OT device manufacturers to build products that are secure by design. The authoring organizations urge device manufacturers to take ownership of the security outcomes of their customers in line with the joint guide Shifting the Balance of Cybersecurity Risk: Principles and Approaches for Secure by Design Software.

• Eliminate default credentials and require strong passwords. The use of default credentials is a top weakness threat actors exploit to gain access to systems.
• Mandate MFA for privileged users. Changes to engineering logic or configurations are safety-impacting events in critical infrastructure. MFA should be available for safety critical components at no additional cost.
• Practice secure by default principles. OT components were initially designed without public internet connectivity in mind. When internet connection becomes necessary, implementing additional security measures is essential to safeguard these systems. Manufacturers should recognize insecure states and promptly inform users so they can make informed risk decisions.
  • Include logging at no additional charge. Change and access control logs allow operators to track safety-impacting events in their critical infrastructure. These logs should be available for no cost and use open standard logging formats.
• Publish Software Bill of Materials (SBOMs). Vulnerabilities in underlying software libraries can affect a wide range of devices. Without an SBOM, it is nearly impossible for a critical infrastructure system owner to measure and mitigate the impact of a vulnerability on their existing systems. See CISA’s SBOM webpage for more information.

Additionally, see CISA’s Secure by Design Alert on how software manufacturers can shield web management interfaces from malicious cyber activity. By using secure by design tactics, software manufacturers can make their product lines secure “out of the box” without requiring customers to spend additional resources making configuration changes, purchasing tiered security software and logs, monitoring, and making routine updates.

For more information on secure by design, see CISA’s Secure by Design webpage.

Validate Security Controls

In addition to applying mitigations, the authoring organizations recommend exercising, testing, and validating your organization’s security program against the threat behaviors mapped to the MITRE ATT&CK Matrix for Enterprise framework in this advisory. The authoring organizations recommend testing your existing security controls inventory to assess how it performs against the ATT&CK techniques described in this advisory.

To start:

1. Select an ATT&CK technique described in this advisory (see Table 1 to Table 10).
2. Align your security technologies against the technique.
3. Test your technologies against the technique.
4. Analyze your detection and prevention technologies’ performance.
5. Repeat the process for all security technologies to obtain a set of comprehensive performance data.
6. Tune your security program, including people, processes, and technologies, based on the data generated by this process.

The authoring organizations recommend continually testing your security program, at scale, in a production environment to ensure optimal performance against the MITRE ATT&CK techniques identified in this advisory.

Resources

Entities requiring additional support for implementing any of the mitigations in this advisory should contact their regional CISA Cybersecurity Advisor for assistance. Key resources organizations should reference include:

• CISA, EPA, NSA, FBI, ASD’s ACSC, Cyber Centre, BSI, NCSC-NL, and NCSC-NZ’s Foundations for OT Cybersecurity: Asset Inventory Guidance for Owners and Operators offers best practices to assist organizations in identifying and prioritizing which assets should be secured and protected.
• CISA, FBI, NSA, EPA, DOE, USDA, FDA, MS-ISAC, Cyber Centre, and NCSC-UK’s guidance on Defending OT Operations Against Ongoing Pro-Russia Hacktivist Activity that can help organizations protect OT systems from pro-Russia hacktivist activity.
• NSA and CISA’s guidance on Control System Defense: Know the Opponent helps organizations defend OT and ICS assets against malicious cyber activity.
• CISA and EPA’s resource page on Water and Wastewater Cybersecurity to help organizations reduce risks posed by malicious cyber actors targeting water and wastewater systems.
  • For additional guidance, see CISA, EPA, and FBI’s fact sheet on Top Cyber Actions for Securing Water Systems.
• The Food and Ag-ISAC’s best practices on Food and Ag Cybersecurity: A Guide for Small & Medium Enterprises provides recommendations to help mitigate against cyber threats.
• DOE and National Association of Regulatory Utility Commissioners Cybersecurity Baselines for Electric Distribution Systems and Distributed Energy (DER) webpage provides resources for state public utility commissions and utilities, as well as DER operators and aggregators to help mitigate cybersecurity risks.

Additional resources that apply to this advisory include:

• EPA’s Cybersecurity for the Water Sector resource page provides organizations with guidance on implementing basic cyber hygiene practices.
• CISA’s Cross-Sector Cybersecurity Performance Goals enables critical infrastructure organizations to reduce the likelihood and impact of known risks and adversary techniques.
• CISA’s Require Strong Passwords webpage supports small and medium-sized businesses mitigating against malicious cyber activity that targets weak passwords.
• CISA, NSA, FBI, EPA, TSA, and international partners’ guidance Secure by Demand: Priority Considerations for Operational Technology Owners and Operators when Selecting Digital Products.
• DOE’s guidance on Cyber-Informed Engineering recommends considering cyber-enabled risks during the conception, design, and development phases when manufacturing physical systems.
• CISA’s Cyber Hygiene Services help enable critical infrastructure organizations to reduce their exposure to threats by taking a proactive approach to monitoring and mitigating attack vectors.
• CISA, NSA, FBI, and international partners’ guidance on Shifting the Balance of Cybersecurity Risk: Principles and Approaches for Secure by Design Software urges software manufacturers to provide customers with products that are safer and more secure.
  • See more information in these Secure by Design Alerts: How Manufacturers Can Protect Customers by Eliminating Default Passwords and How Software Manufacturers Can Shield Web Management Interfaces From Malicious Cyber Activity.

Contact Information

U.S. organizations are encouraged to report suspicious or criminal activity related to information in this advisory to CISA, FBI, and/or NSA:

• Contact CISA via CISA’s 24/7 Operations Center at contact@cisa.dhs.gov or 1-844-Say-CISA (1-844-729-2472) or your local FBI field office. When available, please include the following information regarding the incident: date, time, and location of the incident; type of activity; number of people affected; type of equipment used for the activity; the name of the submitting company or organization; and a designated point of contact.
• For NSA cybersecurity guidance inquiries, contact CybersecurityReports@nsa.gov.

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

Canadian organizations: Report incidents by emailing Cyber Centre at contact@cyber.gc.ca.

New Zealand organizations: Report cyber security incidents to incidents@ncsc.govt.nz or call 04 498 7654.

United Kingdom organizations: Report a significant cyber security incident: report.ncsc.gov.uk (monitored 24 hours) or, for urgent assistance, call 03000 200 973.

Disclaimer

The information in this report is being provided “as is” for informational purposes only. The authoring organizations do not endorse any commercial entity, product, company, or service, including any entities, products, or services linked within this document. Any reference to specific commercial entities, products, processes, or services by service mark, trademark, manufacturer, or otherwise, does not constitute or imply endorsement, recommendation, or favoring by FBI and co-sealers.

Acknowledgements

Schneider Electric, Nozomi Networks, Eversource Energy, Electricity Information Sharing and Analysis Center, Chevron, BP, and Dragos contributed to this advisory.

Version History

December 09, 2025: Initial version.

Appendix A: Targeting Methodologies for Pro-Russia Hacktivist Groups

For further information on targeting methodologies for pro-Russia hacktivist groups, see:

• CISA’s alert Unsophisticated Cyber Threat Actor(s) Targeting Operational Technology;
• The joint fact sheet Primary Mitigations to Reduce Cyber Threats to Operational Technology; and
• CISA’s Russia Cyber Threat webpage.

Appendix B: Additional Designators Used for Cited Groups

The cybersecurity industry and cyber actor groups often use various names to reference actor groups. While not exhaustive, the following are the most notable names used within the cybersecurity community to reference the groups in this advisory.

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

• GRU military unit 74455
  • Sandworm Team
  • Voodoo Bear
  • Seashell Blizzard
  • APT44
• Cyber Army of Russia Reborn (CARR)
  • CyberArmy of Russia
  • Народная CyberАрмия (НКА)
  • People’s CyberArmy of Russia (PCA)
  • Russian CyberArmy Team (RCAT)
• NoName057(16)
  • NoName057(16) Spain
  • NoName057(16) Italy
  • NoName057(16) France
• Z-Pentest
  • Z-Pentest Beograd
  • Z-Pentest Alliance
  • Z-Alliance

Introduction

Note: This advisory was originally published on April 7, 2026, to provide tactics, techniques, and procedures (TTPs) and indicators of compromise (IOCs) related to ongoing cyber exploitation of internet-connected operational technology (OT) devices by Iranian-affiliated advanced persistent threat (APT) actors. The authoring agencies updated this advisory on July 22, 2026, to add new guidance on detecting malicious changes in reusable code modules leveraged within Rockwell Automation PLC programs. It also expands the manufacturer scope to include observed targeting of Schneider Electric, Siemens, and potentially other branded/manufactured PLCs, emphasizing the importance of restricting direct internet access and providing best practice resources for secure deployment.

The Federal Bureau of Investigation (FBI), Cybersecurity and Infrastructure Security Agency (CISA), National Security Agency (NSA), Environmental Protection Agency (EPA), Department of Energy (DOE), United States Cyber Command – Cyber National Mission Force (CNMF), and Department of the Treasury (Treasury) (hereafter referred to as the “authoring agencies”) are urgently warning U.S. organizations of ongoing cyber exploitation of internet-connected OT devices—including PLCs manufactured by Rockwell Automation/Allen-Bradley, Schneider Electric, Siemens, and potentially other manufactured PLCs—across multiple U.S. critical infrastructure sectors. As a result of this activity, organizations from multiple U.S. critical infrastructure sectors experienced disruptions through malicious interactions with PLC project files¹ and the manipulation of data displayed on human machine interface (HMI) and supervisory control and data acquisition (SCADA) displays. In a few cases, this activity caused operational disruption and financial loss.

The authoring agencies assess a group of Iranian-affiliated APT actors is conducting this activity to cause disruptive effects within the United States. The group targeted devices spanning multiple U.S. critical infrastructure sectors, including Government Services and Facilities (to include local municipalities), Water and Wastewater Systems (WWS), and Energy Sectors. The authoring agencies previously reported on similar activity targeting PLCs by CyberAv3ngers (aka Shahid Kaveh Group)—a cyber threat actor affiliated with Iran’s Islamic Revolutionary Guard Corps (IRGC) Cyber Electronic Command (CEC).

Due to the widespread use of these PLCs, and the potential for additional targeting of other branded OT devices across critical infrastructure, the authoring agencies recommend U.S. organizations urgently review the TTPs and IOCs in this advisory for indications of current or historical activity on their networks, and apply the recommendations listed in the Mitigations section of this advisory to reduce the risk of compromise.

If owners and operators discover an affected internet-accessible device in their environment, additional technical measures may be necessary to evaluate the risk of compromise. Please engage your cyber incident response plans and contact the authoring agencies and applicable vendors through existing support channels available to customers and integrators (see Contact Information) to receive support, mitigation, and investigation assistance.

For more information on Iranian malicious cyber activity, see CISA’s Iran Threat Overview and Advisories webpage and the FBI’s Iran Threat and Iran Cyber Threat Overview webpages.

Download the PDF version of this report:

(New, July 22, 2026) For a downloadable copy of July 22, 2026 IOCs, see:

• AA26-097A STIX XML (XML, 29 KB)
• AA26-097A STIX JSON (JSON, 30 KB)

For a downloadable copy of historical April 7, 2026 IOCs, see:

Background Information

Similar Historical Activity Targeting Programmable Logic Controllers

During a similar campaign beginning in November 2023, the IRGC CEC-affiliated cyber threat actors known as "CyberAv3ngers” targeted U.S.-based PLCs and HMIs, causing disruptive effects. Private industry and open sources also refer to this group as Hydro Kitten, Storm-0784, APT Iran, Bauxite, Mr. Soul, Soldiers of Solomon, UNC5691, and the Shahid Kaveh Group. These attacks compromised at least 75 devices, targeting U.S.-based Unitronics PLC devices with an HMI used across multiple critical infrastructure sectors, including the WWS. APT actors developed and deployed custom ladder logic code to these devices, replacing the valid ladder logic with malicious code that continues to be observed to date.

For more information on this group’s activity, see the joint Cybersecurity Advisory IRGC-Affiliated Cyber Actors Exploit PLCs in Multiple Sectors, Including US Water and Wastewater Systems Facilities.

Ongoing Threat Actor Activity Against U.S.-Based Programmable Logic Controllers

The FBI observed Iranian-affiliated APT actors targeting internet-exposed PLCs with the intent to cause disruptions—including maliciously interacting with project files, and manipulating data displayed on HMI and SCADA displays—to U.S. critical infrastructure organizations. Iranian-affiliated APT targeting campaigns against U.S. critical infrastructure have recently escalated, likely in response to hostilities between Iran, and the United States and Israel.

(New, July 22, 2026) At one U.S. victim, the FBI observed the APT actors download a malicious project file to a targeted PLC using configuration software. Analysis indicated the project file retained ladder logic for downstream function but added logic that overrode specific instruction sets responsible for maintaining safe operating parameters in the victim’s environment.

Since at least March 2026, the authoring agencies identified (through engagements with victim organizations) an Iranian-affiliated APT group disrupted the function of PLCs. Organizations across several U.S. critical infrastructure sectors (including Government Services and Facilities, WWS, and Energy Sectors) deployed these PLCs within a wide variety of industrial automation processes. Some of the victims experienced operational disruption and financial loss.

Technical Details

Note: This advisory uses the MITRE ATT&CK^® Matrix for Enterprise framework, version 19. See the MITRE ATT&CK Tactics and Techniques section of this advisory for tables of the threat actors’ activity mapped to MITRE ATT&CK tactics and techniques.

Initial Access

(Updated, July 22, 2026) The authoring agencies observed Iranian-affiliated APT actors using several foreign-based IP addresses to access internet-facing PLCs manufactured by Rockwell Automation/Allen-Bradley, Schneider Electric, Siemens, and potentially other manufactured PLCs [T0883]. The actors used leased, third-party hosted infrastructure and manufacturers’ PLC programming software to connect to misconfigured victim PLCs. Inbound malicious traffic has been observed targeting PLC devices on the following ports: 44818, 2222, 102, and 502, as well as targeting modems on port 22. Targeted devices include:

• Rockwell Automation: CompactLogix and Micro850 PLCs
• Schneider Electric: BMX P34/Modicon M340 PLCs
• Siemens: S7-1200 series PLCs

Command and Control

(Updated, July 22, 2026) The targeting of ports [T0885] associated with other OT vendors’ protocols suggests these actors are opportunistically targeting devices manufactured by companies other than Rockwell Automation/Allen-Bradley, including Schneider Electric and Siemens. In one reported instance, the actors utilized Dropbear Secure Shell (SSH) software on victim modems to enable them to gain remote access through port 22 [T1219].

Exfiltration

(New, July 22, 2026) The authoring agencies observed Iranian-affiliated APT actors using configuration software—such as Rockwell Automation’s Studio 5000 Logix Designer, Schneider Electric’s EcoStruxure Control Expert, and Siemens’ Totally Integrated Automation (TIA) Portal—on leased, third-party hosted infrastructure to exfiltrate device project files from PLC devices to threat-actor-controlled infrastructure [T1041].

Impact

(Updated, July 22, 2026) After the actors extracted device project files, the FBI and CISA identified the modification and deletion of project file logic, to include Add-On Instructions (AOIs) and data manipulation on HMI and SCADA displays [T1565]. Additionally, the changes disabled critical shutdown and alarm logic, allowing systems to enter unsafe conditions without notifying operators of the anomalies.

Note: An AOI is analogous to a “Function Block” or “User Defined Function Block” used in other PLC vendor programs.

Indicators of Compromise

See Table 1 and Table 2 for recent IP addresses used by the Iranian-affiliated APT actors to communicate with PLCs manufactured by Rockwell Automation/Allen-Bradley, Schneider Electric, and Siemens in the United States.

Disclaimer: The FBI observed the threat actors using the IP addresses listed below in the specified time frames. This data is being provided for customers to query against logs for indications of historical targeting by the Iranian-affiliated APT actors. The authoring agencies recommend organizations investigate or vet these IP addresses prior to taking action, such as blocking.

MITRE ATT&CK Tactics and Techniques

See Table 3 to Table 6 for all referenced threat actor tactics and techniques in this advisory. The authoring agencies recommend organizations review historical TTPs for similar Iranian-affiliated cyber actor activity in IRGC-Affiliated Cyber Actors Exploit PLCs in Multiple Sectors, Including US Water and Wastewater Systems Facilities. For assistance with mapping malicious cyber activity to the MITRE ATT&CK framework, see CISA and MITRE ATT&CK’s Best Practices for MITRE ATT&CK Mapping and CISA’s Decider Tool.

Mitigations

The authoring agencies recommend organizations implement the mitigations below to improve your organization’s cybersecurity posture on the basis of the threat actors’ activity. These mitigations align with the Cross-Sector Cybersecurity Performance Goals (CPGs) developed by CISA and the National Institute of Standards and Technology (NIST). The CPGs provide a minimum set of practices and protections that CISA and NIST recommend all organizations implement. CISA and NIST based the CPGs on existing cybersecurity frameworks and guidance to protect against the most common and impactful threats and TTPs. Visit CISA’s CPGs webpage for more information on the CPGs, including additional recommended baseline protections.

Network Defenders

The cyber threat actors accessed PLCs manufactured by Rockwell Automation/Allen-Bradley, Schneider Electric, Siemens, and potentially other branded/manufactured PLCs to cause disruptions to victim systems. To safeguard against this threat and threats to other types of PLCs, the authoring agencies urge organizations to consider the following mitigations.

(Updated, July 22, 2026) In addition to contacting the authoring agencies, organizations and integrators operating PLCs from the manufacturers mentioned in this advisory should review the previously issued guidance to strengthen the security of their OT deployments:

• Rockwell Automation: Contact the Rockwell Automation Product Security Incident Response Team (PSIRT) at PSIRT@rockwellautomation.com for questions regarding this guidance, or to report cyber incidents related to Rockwell Automation products.
  
  • Refer to Rockwell Automation Security Advisory SD1771 for recommended PLC hardening measures and configuration guidance.
• Schneider Electric: Contact the Schneider Electric Corporate Product Cyber Emergency Response Team (CPCERT) at cpcert@se.com for questions regarding this guidance, or to report cyber incidents related to Schneider Electric products.
  
  • Refer to Schneider Electric’s Recommended Cybersecurity Best Practices and Cybersecurity User Guide for Modicon Controller Platform for guidance on securing and configuring PLCs.
• Siemens: Contact Siemens ProductCERT at productcert@siemens.com for questions regarding this guidance, or to report cyber incidents and vulnerabilities related to Siemens products.
  
  • Refer to Siemens Security Bulletin 104599 for a list of security measures to harden PLCs and in-depth configuration guides.
  • Siemens users should review the Cybersecurity for Industry Operational Guidelines and implement defense-in-depth controls within their automation systems.

Immediate steps to prevent the attack:

• Disconnect the PLC from the public-facing internet [CPG 3.S]. Follow the joint guidance Secure connectivity principles for OT to safely allow remote access. Specifically, “remove inbound port exposure,” so the OT system is never directly exposed to the internet or external networks, and to ensure all access is mediated, monitored, and controlled. Do this through a secure gateway (jump host) that brokers the connection.
  
  • Ensure cellular modems, used for remote field connectivity and access, are secured with strong authentication and updated.
  • Enable logs for connected modems and regularly review for suspicious activity to detect intrusions and improve incident response speed.
  • (New, July 22, 2026) To mitigate unauthorized access to OT via cellular modems, organizations should consider implementing isolated architectures, such as private Access Point Name (APN), 5G Public Network Integrated Non-Public Network (PNI-NPN), cellular Software-Defined Wide Area Network (SD-WAN), Zero Trust Network Access (ZTNA), or a site-to-site virtual private network (VPN).
• (New, July 22, 2026) Strictly control network access to PLC devices.
  
  • Configure firewall rules or access control list (ACL) security features on PLCs or programmable controllers to allow only authorized communications between expected control system devices. Block access from unauthorized or threat actor-controlled IP addresses, such as those associated with hosting providers.
• For controllers with a physical mode switch, place the physical mode switch into run position to prevent remote modification. Devices should only be in the program or remote position when updating or downloading software online and immediately switched back to the run position when complete. (See Rockwell Automation’s² System Security Design Guidelines for manufacturer’s instructions.)
  
  • (New, July 22, 2026) Prior to switching the device to run mode, review and validate project files, as changing modes will lock in the current project file downloaded to the device.
• For devices that allow software key switching, enable programming protection in PLC configuration software (S7 TIA Portal) to limit who can modify PLCs remotely. (See Siemens’ Cybersecurity for Industry Operational Guidelines for the manufacturer’s instructions.)

Follow-up steps to strengthen security posture:

• (New, July 22, 2026) Review project files running on PLCs for unauthorized changes. Use vendor-provided integrity checking tools and visually compare the running program to known good logic. Ensure reusable logic and input/output configurations are valid. For Rockwell Automation PLCs listed in the Customer Guidance to Disconnect Devices from the Internet, check the AOIs for any anomalous modifications.
  
  • If restoring from backups, verify the backup does not contain malicious logic before deployment.
  • Review logs and configurations on all connected devices, including modems, HMIs, and workstations, to assess potential lateral movement by threat actors. If it appears the actors connected to additional devices, reimage these devices to remove any potential malicious changes or access tools.
• (New, July 22, 2026) Ensure device passwords are changed from their default and are configured to use complex, unique combinations of letters, numbers, and symbols that are not easily guessable. Implementing robust password practices remains a critical security measure that can help prevent unauthorized access and strengthen the overall security posture of OT devices.
• (New, July 22, 2026) Take defensive measures to minimize the risk of exploitation. Conduct comprehensive impact analysis and risk assessments prior to deploying defensive measures.
• Create and test strong backups of the logic and configurations of PLCs. Store backup files offline and secure the physical removal media to enable fast recovery.
• Implement multifactor authentication (MFA) [CPG 3.F] for access to the OT network from an external network.
• If remote access is required, implement a network proxy, gateway, firewall, and/or VPN in front of the PLC to control network access.
  
  • A VPN or gateway device can enable MFA for remote access even if the PLC does not support MFA. Implement security rules on these higher-level network security mechanisms to prevent the type of repeated and sustained login attempts seen during a brute force attack. When possible, implement a device control list for workstations sending messages or connecting to OT components.
  • Use the device control list to monitor for logon activity for unexpected or unusual access to devices from the internet.
• Keep PLC devices updated with the latest software patches issued by the manufacturer. Use established downtime windows to install patches. Known Exploited Vulnerabilities may need to be prioritized outside a downtime window.
• Configure external and internal firewalls to block traffic using common ports associated with network protocols that are unnecessary for the particular network segment.
• Disable any unused authentication methods, logic, or features, such as default authentication keys and passwords, as well as unused or needed services such as Teletype Network (Telnet), File Transfer Protocol (FTP), Remote Desktop Protocol (RDP), Virtual Network Computing (VNC), and web services.
• Monitor asset management systems for device configuration changes, which can be used to understand expected parameter settings.
• Monitor the content of network traffic for the following:
  
  • Unusual logins to internet-connected devices or unexpected protocols to/from the internet. 
  • Functions of industrial control systems management protocols that change an asset’s operating mode or modify programs.
• (New, July 22, 2026) Ensure service providers are informed of active threats targeting internet-connected PLC devices. Owners and operators should communicate directly with service providers to address risks, especially when remote monitoring or maintenance is involved. Some service providers may rely on internet connectivity essential to monitor and maintain OT/ICS operations but may not be fully aware of active threats.

In addition, the authoring agencies recommend network defenders apply the following mitigations to limit potential adversarial use of common system and network discovery techniques, as well as reduce the impact and risk of compromise by cyber threat actors:

• Reduce risk exposure. CISA offers a range of services at no cost, including scanning and testing, to help organizations reduce exposure to threats via mitigating attack vectors. CISA’s Cyber Hygiene Services can help provide additional review of organizations’ internet-accessible assets. 

Device Manufacturers

Note: The following guidance is general in nature and not specific to any OT vendor. Some of the features, settings, and practices may already be offered by certain vendors. The inclusion of this guidance should not be interpreted as an assertion that vendors referenced do not offer such security features. Also, this advisory is not highlighting a new vulnerability in the identified products, but instead discusses opportunistic targeting. Device manufacturers can make opportunistic attacks more difficult at scale by encouraging more secure behavior by default and in operations, as discussed below. 

Although critical infrastructure organizations using PLC devices can take steps to mitigate the risks, it is ultimately the responsibility of the device manufacturer to build products secured by design and default. The authoring agencies urge device manufacturers to take ownership of their customers’ security outcomes by following the principles in the joint guide Secure by Demand: Priority Considerations for OT Owners and Operators when Selecting Digital Products, primarily:

• Change the manufacturers’ default settings to prevent exposing administrative interfaces to the internet.
• Do not charge additional fees for basic security features needed to operate the product securely.
• Support MFA, including via phishing-resistant methods.

By using secure by design tactics, software manufacturers can make product lines secure “out of the box” without requiring customers to spend additional resources making configuration changes, purchasing tiered security software and logs, monitoring, and making routine updates.

For more information on common misconfigurations and guidance on reducing their prevalence, see joint advisory NSA and CISA Red and Blue Teams Share Top Ten Cybersecurity Misconfigurations. For more information on secure by design, see CISA’s Secure by Design webpage and joint guide.

Validate Security Controls

In addition to applying mitigations, the authoring agencies recommend exercising, testing, and validating your organization's security program against the threat behaviors mapped to the MITRE ATT&CK for Enterprise framework in this advisory. The authoring agencies recommend testing your existing security controls inventory to assess how they perform against the ATT&CK techniques described in this advisory.

To get started:

1. Select an ATT&CK technique described in this advisory (see Table 3 to Table 6).
2. Align your security technologies against the technique.
3. Test your technologies against the technique.
4. Analyze your detection and prevention technologies’ performance.
5. Repeat the process for all security technologies to obtain a set of comprehensive performance data.
6. Tune your security program, including people, processes, and technologies, based on the data generated by this process.

The authoring agencies recommend continually testing your security program, at scale, in a production environment to ensure optimal performance against the ATT&CK techniques identified in this advisory.

Resources

• Authoring Agencies: IRGC-Affiliated Cyber Actors Exploit PLCs in Multiple Sectors, Including US Water and Wastewater Systems Facilities
• CISA: Bulletproof Defense: Mitigating Risks From Bulletproof Hosting Providers
• EPA: Cybersecurity for the Water Sector
• CISA: Water and Wastewater Cybersecurity
• CISA: Exploitation of Unitronics PLCs used in Water and Wastewater Systems
• CISA: Iran Threat Overview and Advisories
• FBI: The Iran Threat and Cyber Threat Overview: Iran
• CISA, MITRE: Best Practices for MITRE ATT&CK Mapping
• CISA: Decider Tool
• CISA: Cross-Sector Cybersecurity Performance Goals 2.0
• CISA: No-Cost Cybersecurity Services and Tools
• CISA: Secure by Demand: Priority Considerations for OT Owners and Operators when Selecting Digital Products
• NSA, CISA: NSA and CISA Red and Blue Teams Share Top Ten Cybersecurity Misconfigurations
• CISA: Secure by Design
• FBI, CISA: Primary Mitigations to Reduce Cyber Threats to Operational Technology
• United Kingdom National Cyber Security Centre: Secure connectivity principles for operational technology

Contact Information

U.S. organizations are encouraged to report suspicious or criminal activity related to information in this advisory to CISA, the FBI, and/or NSA:

• Contact CISA via CISA’s 24/7 Operations Center at contact@cisa.dhs.gov or 1-844-Say-CISA (1-844-729-2472). File a claim with FBI’s Internet Crime Complaint Center (IC3) or contact your local FBI field office. When available, please include the following information regarding the incident: 
  • Date, time, and location of the incident;
  • Type of activity;
  • Number of people affected;
  • Type of equipment used for the activity; and
  • Name of the submitting company or organization, and a designated point of contact.
• For NSA cybersecurity guidance inquiries, contact CybersecurityReports@nsa.gov.
• Entities required to report incidents to DOE should follow established reporting requirements, as appropriate. For other energy sector inquiries, contact EnergySRMA@hq.doe.gov.
• Contact the Rockwell Automation PSIRT for questions regarding their guidance or for reporting cyber incidents related to Rockwell Automation products at PSIRT@rockwellautomation.com.
• Contact the Schneider Electric CPCERT at cpcert@se.com for questions regarding this guidance, or to report cyber incidents related to Schneider Electric products.
• Contact Siemens ProductCERT for up-to-date information about the security of Siemens products or to report cybersecurity vulnerabilities at productcert@siemens.com. For support with increasing the security of installed Siemens PLCs, contact Siemens Industrial Cybersecurity Services at services.automation@siemens.com. See Siemens ProductCERT and Siemens CERT for more information.

Disclaimer

The information in this report is being provided “as is” for informational purposes only. CISA and the authoring agencies do not endorse any commercial entity, product, company, or service, including any entities, products, or services linked within this document. Any reference to specific commercial entities, products, processes, or services by service mark, trademark, manufacturer, or otherwise, does not constitute or imply endorsement, recommendation, or favoring by CISA and the authoring agencies.

Version History

April 7, 2026: Initial version.

July 22, 2026: Update includes new guidance on detecting malicious activity, expanded scope of observed targeting, and best practices for secure PLCs deployment.

Notes

¹Project file refers to the software file that contains ladder logic and configuration settings. On Rockwell Automation devices, it is referred to as an .ACD file.

² See CompactLogix 5370 Controllers (Chapter 5: “Select the Operating Mode of the Controller”) for more information on functions available for the switch.

• James enabled adaptive autofill in Nightly for testing, which we believe should provide better results in the URL bar when doing autocomplete!
• Jack updated the illustrations shown on some of our error pages to match the latest approved designs, giving users more polished artwork when the browser encounters connection or security errors!

• Controls for the Memories feature can now be set during Smart Window onboarding

• We’ve disabled the CSS filter implicitly applied to WebExtension pageAction SVG icons across all release channels starting in Firefox 152, completing the deprecation
  • NOTE: The blog post published at WebExtensions API changes in Firefox 149-152 provides to extensions developers more details about this deprecation and links to the related MDN docs.

Friends of the Firefox team

Resolved bugs (excluding employees)

Script to find new contributors from bug list

Volunteers that fixed more than one bug

• Amin Amir
• Pranjali Srivastava
• Sam Johnson

New contributors (🌟 = first patch)

•  🌟:23rd: Regression: The new swipe-to-navigation indicator stucks for a moment, when deciding not to navigate the other page
• 🌟Akeem Omosanya: Remove commented-out code in SearchService.sys.mjs
• Amin Amir:
  • 🌟Fix browsingContext.sys.mjs to assign to #contextCreatedHandled instead of contextCreatedHandled
  • Fix missing WITHOUT ROWID SQLite performance optimization in SERPCategorization.sys.mjs
• 🌟Sahaj: Suggest the default target language for translation after changing the detected source language
• 🌟JIANG Zhirui: Breakpad build failed on Windows using VS2026 due to removal of stdext
•  John Iweh: Add “Open in New Tab” and “Open in New Container Tab” options to the context menu for Tabs from Other Devices
• Jak: Bookmarks and History – should respect the “When you open a link, image or media in a new tab, switch to it immediately” setting
• 🌟Andy [:rgbcmy]: Autoplayed next video should also be PIP
•  konyhéa: “Escape” key should collapse the expanded on hover sidebar launcher even if hover is still active.
•  Pranjali Srivastava:
  • Remove icon property from sidebar extensions
  • Show language-agnostic SelectTranslations context menu item when the source and target languages are the same

Project Updates

Add-ons / Web Extensions

Addon Manager & about:addons

• Fixed long-standing regression on the autocomplete and datalist popups for extension inline options pages on about:addons (introduced in Firefox 68 by Bug 1532724, fix shipping in Firefox 152) – Bug 1595158

WebExtensions Framework

• Fixed access to web-accessible resources declared with <all_urls> from sandboxed documents (null-principal URLs), restoring extension redirects from the context-menu search flow, starting in Firefox 152 – Bug 2033905

WebExtension APIs

• Added exhaustive test coverage for tabs.move() against additional edge cases related to split-view tabs – Bug 2029092

DevTools

• Andreas Farre improved the Session History tab in the Application panel (still behind devtools.application.sessionHistory.enabled)
  • added support for remote debugging (#2014064, #2016121)
  • made sure that calls to History.replaceState are reflected in the UI (#2037359)
• Julian Descottes [:jdescottes] fixed the most frequent DevTools crash we were observing in Telemetry, adding a guard against IDBTransaction errors when retrieving breakpoints in the Debugger (#2030260)
• Nicolas Chevobbe [:nchevobbe] fixed the image preview tooltip for relative URLs images in constructed stylesheet (#2035503)
• Julian Descottes [:jdescottes] reduced the overhead we had because of network requests monitoring by only decoding response content when the user actually want to see the response (#2026228)

WebDriver

• Amin Amir cleaned up an incorrect variable assignment in our browsingContext module.
• Logan Rosen updated stale references and broken links in our documentation about Marionette.
• Sameem improved the Marionette and WebDriver BiDi screenshot commands to enforce maximum allowed dimensions.
• Leo McArdle fixed the regression in the “log.entryAdded” event, which lacked an error message in the “text” field for the messages of type “error”.
• Henrik Skupin fixed an issue in Marionette where WebDriver:Navigate and WebDriver:Refresh did not handle errors when the underlying navigation failed.
• Henrik Skupin improved geckodriver to detect an early Firefox exit during startup on Android, avoiding up to 60 seconds of unnecessary connection attempts.
• Henrik Skupin updated the geckodriver CI build job to produce a universal macOS binary supporting both x64 and aarch64.

Lint, Docs and Workflow

• Sylvestre ported some linters (e.g. file-whitespace, test-manifest-toml, license, file-perm, rejected-words & more) to Rust to help improve the runtime of the code review bot.
• Dale has been working on migration to moz-src for customkeys, dom/quota and odom/geolocation
  • https://arewemozsrcyet.com/

New Tab Page

• We did our first region-specific trainhop on May 11th (just 15% of the US), and turned on HNT Nova (and sometimes Widgets) for those clients to get some advance-data of its behaviour in the wild! A note that HNT Nova gets turned on for everybody when Firefox 151 ships on May 19th.
  • We’ll be launching a similar experiment in the DE, probably on May 12th, also at 15% population.
• Most of the team is heads down building out a sports-tracking widget, attempting to get that ready in time to be generally available for the upcoming World Cup event.
• Dre landed a new world clock widget, which is currently off by default, but pretty snazzy!

Search and Urlbar

• Nova (URL Bar Design Refresh)
  • Drew and Daisuke continued their work on Nova styling for the Address bar (input and view).
• Search and Suggest
  • Drew finalized two bugs for World Cup and sports suggestions, which were landed and uplifted: one to update the localization string for scheduled games and another to show both teams’ icons in suggestions. Drew also landed and uplifted a fix for rich search suggestion icons being forced into a square aspect ratio.
  • Standard8 updated Ecosia favicons to the latest branding, including QA testing and publishing.
• Settings Redesign (SRD)
  • Stephanie landed a test to ensure search suggestion settings are hidden when quicksuggest is disabled, as well as a patch to resolve TypeScript issues in search.mjs, and is adding test coverage to confirm removed search engines are not displayed in the default engines dropdown.
• General URL Bar and Component Updates
  • Daisuke landed implementation of the context menu on URL bar results, and a fix to show the loading URL in the URL bar when starting up with a homepage.
    • Marco is working on several tasks, including a PDF download / focus stealing issue and allowing arrays to be bound in Sqlite.sys.mjs. Marco also worked on fixes related to Places, such as avoiding replacing the favicons database if it is not corrupt.
  • Standard8 finalized the URL bar test manifest split. Standard8 also upgraded us to TypeScript 6.
  • Moritz landed a fix for URL bar abandonment telemetry being recorded when clicking an engine in the unified search button popup (Bug 2032973), which was also uplifted. Moritz also simplified search mode switcher item activation in tests, and made it so that the unified search button popup closes when installing an open search engine.

Smart Window

• natural language starting with tab close/undo 2035343 with expandable action log 2031508

• assistant rendering feedback up/down 2032994 and markdown table 2027029
• nova styling blur 2027877 and suggestions 2026823
• accessibility screen reader 2028676 and keyboard focus 2037565
• optimize conversation starters extra requests 2030005 and caching 2033430

Storybook/Reusable Components/Acorn Design System

• Nova token updates occasionally, focused on SRD

UX Fundamentals

• Added support for the “SEC_ERROR_CA_CERT_INVALID” certificate error to the Felt Privacy error pages. – 2035942

Settings Redesign

• Settings redesign is being tested and will hopefully go out in Firefox 152!

Bots (and privacy-preserving browsers) not welcome

Browse a news site in a private window. Shop at a major retailer with a VPN. Visit a video streaming platform with anti-fingerprinting defenses tuned up. You’ll see the same responses: registration walls, block pages, and endless CAPTCHAs. The message is clear: if we think you might be a bot, you’re not welcome.

Websites have valid reasons for wanting to block bots. Bots enable volumetric abuse, abuse that wouldn’t otherwise be feasible if they had to be carried out by humans.  For example: SEO comment spam, credential stuffing and DDoSing. Consequently many sites employ dedicated anti-abuse tooling which aims to keep the bots out whilst minimizing friction for human visitors.

Unfortunately, that tooling is increasingly failing at both tasks. Browser privacy protections are dismantling the passive signals that anti-abuse systems depended on to identify and distinguish visitors. Meanwhile advances in generative AI have rendered CAPTCHAs ineffective: bots now solve them faster and more reliably than humans.

Many sites are switching to more invasive mechanisms and now ask visitors to disclose identifying information, e.g. an email address, a federated login or disabling their VPN. This means greater friction for users, since providing these details on a first visit takes time. It also compromises their privacy, since these details enable the same kinds of cross-site tracking that browser privacy protections were intended to mitigate.

This leaves users with a dilemma. The more effectively they protect their privacy, the harder it is for websites to distinguish them from bots and the worse the treatment they receive. Website operators are also suffering. The additional friction they inflict upon well-behaved visitors harms their site, but many are willing to pay the costs if it mitigates volumetric abuse.

Browser-based AI agents make this tension more acute. Sites may want to allow agents which are acting on behalf of individual users while blocking agents engaged in volumetric abuse. However, with no effective mechanisms to distinguish the two, websites are opting to block both. That hurts users, who should be free to choose the user agent they use to access the web; it hurts new browsers and agents, which struggle to interoperate; and it hurts sites, which lose legitimate visitors.

The consequence is that the web gets worse for everyone. Users get more friction or less privacy or both. Website operators see more volumetric abuse and the friction they add drives away users who would otherwise want to consume their content or services. New user agents struggle to access the same content as conventional browsers.

The Costs of Convenient Solutions

Some large ecosystem players have put forward solutions that leverage their control of the dominant operating systems and their deep integration with consumer hardware. These rely on device attestation: identifiers and privileged code baked into devices at the hardware level, which let manufacturers prove what software is running on a user’s device. Exposing this functionality to the web means attesting to sites that the user is running approved software with trusted hardware and therefore isn’t a bot. There have been two substantive proposals.

Google’s Web Environment Integrity, abandoned in 2023, was the blunt version. It attested to the user agent itself, as well as the operating system and device in use. Users would have lost control in two ways: once to the attester, which would decide which operating systems and devices could be blessed, and again to the website, which would decide which software to accept. If sites had adopted allow-lists of approved user agents, building a new browser would have become virtually impossible, and sites could have withdrawn access from any user agent they chose.

Apple’s Private Access Tokens, deployed across their ecosystem in 2022, have more subtle issues. Built on the Privacy Pass protocol standardized at the IETF, they get a lot right: a user receives a renewed, limited batch of one-time tokens that can be presented to websites without linking their visits together. This provides privacy for users and has shown rate limits to be an effective tool for sites – both points we’ll return to later in this post.

However, Private Access Tokens rely on device attestation, requiring that the hardware manufacturer be in overall control of the user’s device. Presenting a PAT tells a website you are locked into Apple’s rules for what counts as acceptable software. Due to PAT’s technical design^[1], there’s no way to open the system to other sources of scarcity without compromising the system’s privacy properties, meaning that if more widely deployed, access to the web would become tied to having bought expensive hardware from a small, hard to change set of vendors.

Both approaches are ultimately hostile to users and to the openness of the web. Both are premised on parts of a user’s device that sit within the manufacturer’s control and beyond the user’s own. Were they widely deployed, the web would become just another walled garden with centralized gatekeepers controlling acceptable hardware, operating systems and software. As convenient as these solutions are for the players who already dominate the ecosystem, we think there’s a better path.

A Better Path Forward

Bots’ harms arise from their ability to operate beyond human scale. For sites to prevent volumetric abuse they don’t actually need to know the user’s identity or receive cryptographic proof that they’re running approved software. If sites knew their visitors were restricted to a rate limit set by a site, that would be enough.  

Rate limits only make sense if they’re tied to something scarce; something an attacker can’t cheaply replicate to evade the limit. Without anchoring to a scarce resource, like the trusted hardware used in Private Access Tokens, attackers can generate as many fresh identities as they need to bypass the rate limit.

However, hardware is just one option for scarcity. Anything a user already has that an attacker can’t trivially spin up at scale will work: email addresses and phone numbers are naturally scarce. A paid subscription costs an attacker the same as a real user.  Even maintaining an account on a free service requires some non-trivial work.

What if we could use these scarce signals across the web? We could build an open ecosystem with many parties offering scarcity signals, each site choosing which to accept. By opening up who can provide a signal, and letting sites choose which to accept, we can avoid transferring control to device manufacturers and the resulting harms.

As a concrete example of who might be well positioned to provide such a signal, we can consider VPN providers acting as a subscription service. Sites routinely block VPN users indiscriminately, whether through a deliberate policy choice or through an indirect consequence of rate limiting visitors per IP address. But a VPN subscription is a perfect source of scarcity. If the VPN provider could vouch for its users so that sites could rate limit each user individually – then users would be able to browse the web with less friction and without giving up their VPN usage.

The catch is that building a system that can enable this on the open web whilst maintaining user’s privacy is genuinely difficult. It requires that we take information from one site — that this user holds some scarce thing — and expose it to other sites so that they can use that as the basis for their rate limiting. Letting one site verify a signal from another is the sort of information flow that privacy-preserving browsers have spent the last decade locking down to prevent cross-site tracking.

Our goal would be that no more than the minimum information gets through: a single bit communicating whether the user is below the rate limit set by the site. Leaking anything more – like the source of the scarcity that the rate limit is anchored to – would be unacceptable. Enabling a new cross-site information flow might feel like compromising privacy to gain better access, but reality is more nuanced. If a new system moves sites away from demanding that visitors be identifiable (whether through fingerprinting or login forms), it can be a win for both privacy and access.

The Foundations

The good news is that the cryptographic foundations for a privacy preserving approach already exist. The Privacy Pass protocol, originally developed in 2018 to reduce the friction of Cloudflare CAPTCHAs for Tor users, introduced the core primitive: a token that is unlinkable between issuance and redemption. You prove something to an issuer (e.g. by solving a CAPTCHA), receive some tokens, and later present a token to a website. The website can verify the token is legitimate, but can’t link it to the user it was issued to.

Figure 1: In Privacy Pass, a CAPTCHA provider can issue tokens to a client which can then be used to bypass challenges for future site visits. Even if the CAPTCHA provider and sites collude, they can’t use the tokens to identify the user or their browsing history.

Privacy Pass has gone on to be successfully deployed in systems where the issuer and verifier have a prior trust relationship: Apple uses it to authenticate users of Private Cloud Compute and Private Relay without linking their activity to their identity, Chrome uses it for two-hop IP protection, and Kagi uses it to provide private search. These deployments work in part because a small number of parties have agreed in advance on who issues tokens and who accepts them.

Applying this approach to an open system where any site can act as an issuer brings real challenges. Firstly, even though tokens are unlinkable, knowing a user has access to a specific issuer is a privacy leak on its own, because you can infer that the user meets the relevant issuance criteria. If one site can learn that you have a token from another site, that reveals that you have been to that site, which can be a major privacy problem. This compounds if sites can learn the set of issuers you have visited, since it becomes a fingerprint which can be used to identify you.

Generic techniques exist for proving a statement in zero knowledge: we can prove that a client has a token from a set of acceptable issuers without revealing which specific issuer it is. We’ll call this issuer blinding. The generic approach is often slow, but bespoke approaches tailored to the underlying cryptography can improve this considerably.

Another challenge is how sites using rate limits decide who to trust to issue tokens. If an issuer misbehaves then the site’s rate limits become ineffective, enabling volumetric abuse. However, if we need to prevent the site from learning which issuers a user has access to, the site is only going to know that one of its trusted issuers was used, not which one. This makes mistakes or misbehaviour by an issuer difficult to detect, and makes it hard for sites to evaluate new issuers. Solving this challenge is essential for openness. Without adequate information, sites are likely to lean towards conservative issuer selection. That could lead to less choice between Anchors, which in turn could lead to a new form of gatekeeper being created. 

To solve this, sites at least need a way to calculate an aggregate score for each issuer they use. This should roughly correspond to how much of the traffic it considers abusive to have come from users using that particular issuer. Mozilla has long invested in systems like Prio which use multiparty computation (MPC) to protect user privacy whilst enabling aggregate measurements of system behaviour.

Privacy Pass also struggles to handle dynamic adjustments to rate limits. Once tokens have been issued, they’re difficult to invalidate without either revoking all active tokens or risking attacks which can compromise the privacy of users. It’s also beneficial if sites can adjust rate limits on a per client basis, for example by increasing rate limits where they become more confident the client is benign and withdrawing access when abuse is detected.

Anonymous Credit Tokens offer a useful building block to solve this problem. Conventional Privacy Pass schemes rely on issuing a bucket of tokens but ACT works differently by enabling the use of a credential with state. For example, an ACT credential can hold an internal counter. When the credential is presented, the site can check the counter is over some threshold and mutate it, increasing or decreasing the counter whenever the site’s perception of the holder has improved or worsened. Critically, the exact value is never leaked to the site, preventing the site from tracking the holder and ensuring successive presentations of the same credential can’t be linked.

Putting it together

So how can we combine these techniques to build a system which can enable privacy-preserving rate limiting on the open web? In May 2026, we participated in a W3C CG Meeting in collaboration with Cloudflare, Chrome and other web stakeholders in which we started sketching out a design we’re calling PACT – Private Access Control Tokens.

Rate limits need a starting point, a source of scarcity to anchor on. We’ll call an entity that provides such a source an Anchor. To a user who meets the Anchor’s criteria, like having a subscription, an account in good standing, or a verified phone number, an Anchor issues a batch of Endorsement tokens, following the Privacy Pass model. In practice, Anchors could be any website which has access to this kind of signal. An Endorsement conveys scarcity to other sites.

That’s enough for a simple system where access is either granted or denied. But as we discussed earlier, we also want the ability to increase access where a visitor behaves benignly and decrease it where they don’t. The state needed to enforce a rate limit can’t live in the Endorsement, because Endorsements cross trust boundaries between unrelated sites. We need a second object that can hold that state, scoped to the party that maintains it.

We’ll call that the party that handles rate limiting for a site a Moderator and the stateful object a Credential. A Credential is specific to a Moderator and, unlike endorsements, we limit each site to nominating a single Moderator. In the common case the site itself plays the Moderator role, so there’s no new entity or trust boundary. A Moderator can also be a third-party service shared across many sites, allowing those sites to cooperatively share a rate limit. 

In the terminology of the previous section, the Anchor is the issuer of Endorsements, and the Moderator both verifies Endorsements and issues Credentials. A Moderator manages rate-limit policy: it decides which Anchors it trusts, accepts their Endorsements, and issues a Credential in return.

Figure 2: (1) Clients acquire Endorsements from Anchors in the course of normal browsing to sites they have relationships with. (2) Clients can exchange Endorsements for a stateful Credential from a Moderator. (3) Credentials can be used to access sites which use that Moderator. Credentials can be updated over time.

Directly revealing which Anchor backed an Endorsement would leak a lot of information about the user. The issuer blinding techniques from the previous section solve this: when an Endorsement is redeemed, the Moderator only learns that it came from one of the Anchors it trusts, but not which one.

When a Moderator covers more than one site, we let Credentials be presented across all of them but partition cookies and storage as we would for any other third party site. The unlinkability of Credential presentations keeps this from creating a new cross-site identifier. The benefit is that good behaviour on one site improves access on every site the Moderator covers, and bad behaviour cuts it everywhere. Websites can already build the same capability with a shared account system, so this doesn’t create a new way to lock users out, but it does provide a new way to grant access without requiring users to give up their privacy.

Enabling Moderators that cover many sites carries a centralisation risk, similar to the concentration we see today in anti-abuse providers. The mitigation is that the choice of Moderator stays with each site, and the choice of trusted Anchors stays with each Moderator. This can’t reverse the centralisation pressure the web already faces, but it ensures this system won’t lead to additional lock-in: a new Anchor or a new Moderator can be adopted without coordinating with a dominant vendor.

The system then has three flows. First, the user receives Endorsements from an Anchor in the course of normal interaction, based on the Anchor’s positive view of the user. This is a relatively rare operation for any given user and Anchor. After all, as our source of scarcity, Endorsements should not be too easy to accumulate.

Figure 3: In the course of normal browsing, clients browse to websites they have a relationship with. These sites can act as Anchors by issuing Endorsements to clients.

Second, when the user arrives at a site that works with a Moderator, the browser spends an Endorsement from an Anchor the Moderator trusts and receives a Credential in return. The presentation hides which Anchor was used, and neither the Anchor nor the Moderator can trace the Endorsement back to where it was issued. The Moderator decides what initial balance the Credential starts with. If the user has no Endorsements from suitable Anchors at all, existing mechanisms (CAPTCHAs, account creation, federated login) could be used to bootstrap a Credential the same way, so the system degrades to today’s experience rather than locking the user out.

Figure 4: When the client browses to a site, it can prompt the client for a Credential from the Moderator it uses. If the Client doesn’t have a suitable Credential, but does have a suitable Endorsement, it can exchange it for a Credential with the Moderator. In practice, the Moderator and the Site might be the same server.  

Third, as the user browses, the browser presents the Credential and the Moderator updates the internal state of the Credential. The Moderator can reward behaviour that looks benign and penalize suspicious activity, but can’t track the use of the Credential or identify it if it’s used on other sites the Moderator covers. Revocation falls out of the same mechanism: a Moderator can refuse to return an updated Credential. 

Figure 5: The Client can present the Credential on sites which use the matching Moderator. Sites can check if the Credential is in good standing. The sites can then adjust the access the Credential has in response to behaviour. E.g. increasing it when they gain confidence in the client or reducing it in response to malicious behaviour.

In practice, all of this would happen transparently to the user through a WebAPI that sites acting as Anchors or Moderators would call from JavaScript. In an ideal ecosystem, users would accumulate Endorsements through normal browsing, just by virtue of the sites they already visit, and the rest of the flow would happen in the background as they move around the web, leaving users with meaningfully less friction.

AI agents acting on behalf of a user slot into the same flow. An agent can carry its user’s Credentials, in which case the user remains accountable for how the agent behaves. Sites would not need to grant any more access than they would to the user themselves. Alternatively, the operator of an agent can run its own Anchor and vouch for its agents the way other Anchors vouch for human users. Sites retain control over which Anchors they accept, so they can choose how to treat agent traffic without needing a separate detection mechanism.

Several mechanisms combine to keep the information about a user that flows out close to a single bit. Cryptographic unlinkability ensures successive Credential presentations cannot be tied to each other or to the original issuance, so a user’s visits cannot be joined into a history. Each site is bound to a single Moderator, so the set of Moderators a user has Credentials with never becomes a cross-site fingerprint. The Anchor-to-Credential exchange happens in an isolated browsing context, so during ordinary browsing the only thing the site or its Moderator ever observes is a Credential presentation: the site only learns if the user has a valid Credential below the rate limit, or nothing. When the Moderator updates a Credential, it adjusts the credentials state without learning what it is.

The additional privacy given to users from Issuer blinding makes participating in the system more challenging for Moderators. Because the Moderator can’t see which Anchor backed a Credential at issuance, it can’t give a Credential from a strong Anchor more access than one from a weak Anchor: doing so would itself leak which Anchor was used. The initial access has to be uniform across the Moderator’s whole pool of Anchors, which in practice means setting it at the strength of the weakest. However, this is only relevant for that initial access, the Moderator can update credentials according to the holder’s behavior, enabling Credential’s to accrue access over time.

Building an open ecosystem also requires that sites can make effective decisions about the Anchors they choose to trust. Multiparty computation systems like Prio enable aggregate scoring without compromising privacy. When users present Credentials, they can provide an encrypted share which identifies the anchor they used and can be privately aggregated to compute the quality of an issuer.

Next Steps

We think the architecture we’ve sketched for PACT has the right shape, but many of the details still need to be worked out and the entire system needs rigorous privacy and security analysis.

We want to do that work in the open. The IETF is the natural venue for the cryptographic protocols underneath, and the W3C for the WebAPI surface that sits on top. We’ll be bringing draft specifications to these bodies as soon as they’re ready, and we welcome collaborators from across the ecosystem: browser vendors, site operators, anti-abuse providers, and the cryptography community.

If successful, we think we can provide a system which will keep the web open and private, while still giving sites the rate-limiting signal they need.

Acknowledgements

The ideas described here are the result of collaboration and conversations with many people, including: Watson Ladd, Thibault Meunier, Michele Orrù, Trevor Perrin, Eric Rescorla, Samuel Schlesinger, Martin Thomson, Eric Trouton, Benjamin Vandersloot & Cathie Yun. 

The post PACT: Anonymous Credentials for the Web appeared first on Mozilla Hacks - the Web developer blog.

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.

- 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

RAM prices have exploded over the past two years, but a little-known Chinese company may finally be changing that. While Samsung, SK hynix, and Micron shifted their factories toward high-profit AI memory, everyday DDR5 RAM became harder to find and dramatically more expensive. Now, China's CXMT is stepping in with a massive expansion of DDR5 production, potentially bringing competition back to one of the world's most concentrated semiconductor markets. In this video, we break down why RAM prices skyrocketed, how AI created the global memory shortage, CXMT's rapid rise, its new DDR5 and LPDDR5X lineup, the company's record-breaking IPO, HBM ambitions, and whether this could finally make PC memory affordable again.

Could CXMT become the fourth memory giant—and finally end the RAM shortage?

#CXMT #DDR5 #RAM #Semiconductors #AI #China #Technology #PC

Thanks to Micro Center for sponsoring this video! Shop Back to School Tech Deals: https://micro.center/baef21

Sign up for a FREE 128 gig Flash Drive at Micro Center Austin, TX: https://micro.center/fdf6b0
Sign up for a FREE 128 gig Flash Drive at Micro Center Columbus, OH: https://micro.center/4f8a89

Visit Micro Center News: https://micro.center/a34dc2

Shopping is getting harder all the time. Virtual aisles filled with crap and fake reviews, but thankfully new tech is....er... making things worse... If you've been burned by things getting more expensive or changing price this video is for you.

Leave a reply with your requests for future episodes.

► SHOP OUR PRODUCTS: https://lttstore.com
► GET A VPN: https://www.piavpn.com/TechQuickie
► GET EXCLUSIVE CONTENT ON FLOATPLANE: https://lmg.gg/lttfloatplane
► SPONSORS, AFFILIATES, AND PARTNERS: https://lmg.gg/partners

VentureBeat Research measured the five controls an enterprise has to build before it can trust an agent: identity, evaluation, cost telemetry, the context layer, and orchestration. Identity governs which agent is allowed to do what, under whose credentials. Evaluation determines whether the agent's work is any good. Cost telemetry tracks what each agent costs to run. The context layer supplies the business data and definitions agents draw on when they answer. And the orchestration control plane coordinates multi-step agent work. Each of our five reports measures one of those controls.

Most deployed "agents" are chatbots wearing the label. Seventy-one percent of enterprises said a quarter or fewer of their deployed "agents" can complete multi-step work on their own; only 10% said true agents are the majority of what they run. These respondents are positioned to know: 81% recommend or decide AI purchases at their companies. A single-prompt chatbot with a human reading every answer needs none of the controls the other four reports measure. A true multi-step agent needs all of them — and most enterprises can't say which one they've deployed. (Full findings: Agentic Orchestration report.)

Autonomy is outrunning trust in the evaluations that gate it. Two-thirds of enterprises either already allow an agent to push a code or system change to production on automated evaluation results alone, with no human review, or are actively engineering toward that within 12 months. Only 5% fully trust the evaluations that would make that call — and half of enterprises shipped an agent that passed internal evaluations and then caused a customer-facing failure in the past year. Before removing human review from any workflow, test evaluations against production outcomes rather than internal benchmarks. (Full findings: Agent Reliability & Evals report.)

Companies that let agents share credentials get hit more often. Sixty-nine percent of companies let at least some of their agents share credentials — multiple agents operating under one API key or service account. Organizations that allow credential sharing anywhere experienced a security incident or near-miss at a 63.5% rate (47 of 74), against 40.9% (nine of 22) at companies where every agent has its own scoped identity. The fix is scoped identity for every agent, starting with the ones that touch production systems. (Full findings: Agentic Security & Identity report.)

The most expensive hardware in the building runs at half capacity or less. More than eight in 10 enterprises that run their own GPUs reported utilization of 50% or less, and only 44% rigorously track what their AI compute actually costs and returns. The number worth chasing first isn't more GPUs — it's the utilization and per-workload cost of the ones already running. (Full findings: AI Infrastructure & Compute report.)

Agents answer confidently from data nobody governs. Fifty-seven percent of enterprises traced a confident, wrong agent answer in the past six months to their own missing or inconsistent business context — wrong metrics, stale definitions, absent documents — and most saw it happen more than once. Governing the definitions agents answer from — metrics and entities first — has to come before scaling the agents that depend on them. (Full findings: Context Layers / RAG report.)

No layer has an entrenched incumbent: The defaults today are the built-in tools that ship with the big AI platforms enterprises already use. Switching intent runs highest in orchestration itself, where 68% plan to adopt, add, or replace platforms within 12 months and 34% within the quarter. Our surveys did not ask which direction that money moves — toward the platforms' built-in tools or toward the specialists challenging them — and that open question is the next four quarters of this market.

About this research

VentureBeat Research fielded five parallel surveys in June 2026 under its VB Pulse program: Agentic Orchestration (101 respondents), Agent Reliability & Evals (157), Agentic Security & Identity (107), AI Infrastructure & Compute (107), and Context Layers / RAG (101) — 573 qualified respondents in total, all at organizations with 100 or more employees. Samples are self-selected, and some findings should be read directionally; each report carries its full methodology note. What the pattern supports more strongly than any single percentage is the direction: every survey, independently, points the same way. VentureBeat produces both this research and VB Transform, the conference where these reports debuted.

The model, available immediately on all of Anthropic's platforms, is priced at $5 per million input tokens and $25 per million output tokens, unchanged from its predecessor, Opus 4.8. It becomes the new default model on Claude Max, Anthropic's premium consumer tier, and the strongest model available on Claude Pro.

The positioning is deliberate. Anthropic is not claiming Opus 5 is its smartest model — that distinction still belongs to Fable 5, and rival systems retain an edge in certain domains. Instead, the company is making a subtler argument that may matter more to enterprise buyers: that the most economically important AI work happens in a middle band of difficulty, where near-frontier intelligence delivered efficiently and cheaply beats frontier intelligence delivered expensively.

"Opus 5 as your daily driver, the model you hand complex work to and review when it's done," an Anthropic spokesperson said in an interview with VentureBeat, describing how the company's lineup now stratifies. "Fable 5 for your most ambitious work, the days-long autonomous projects nothing could take on before... Sonnet 5 for work you run at scale, where speed and cost per call decide what ships. Haiku 4.5 for subagents and instant answers."

How Claude Opus 5 benchmark results stack up against Fable 5 and rival AI models

On paper, the results are striking. Anthropic says Opus 5 sets new state-of-the-art marks on coding and knowledge-work evaluations including Frontier-Bench and GDPval-AA. On Frontier-Bench v0.1, an agentic terminal coding benchmark, Opus 5 scores 43.3 percent — more than double Opus 4.8's 18.7 percent and well ahead of Fable 5's 33.7 percent — at a lower cost per task, according to the company. On ARC-AGI 3, an evaluation of novel problem-solving, Anthropic reports Opus 5 scored three times as high as the next best model. On OSWorld 2.0, a computer-use benchmark, the company says the model surpasses Fable 5's best result at just over a third of the cost.

The numbers come with honest caveats that are themselves notable in an industry prone to superlatives. Anthropic acknowledges Opus 5 remains behind Mythos 5, a competing model, on cybersecurity tasks and biology research, and an OpenAI-family model still leads on one agentic coding benchmark.

The more revealing caveat came from Anthropic itself, when asked where Opus 5 still falls short of Fable 5. The spokesperson's answer amounted to a candid admission about what benchmarks do and don't capture.

"The evals where Opus 5 wins are bounded tasks with a specific outcome, which is where it's strongest. What those evals don't measure is duration," the spokesperson told VentureBeat. "One way to put it: Opus 5 is the best tool for the jobs benchmarks can see, and Fable 5 is what you reach for when the job outruns the benchmark."

Fable 5, by contrast, "is for the longest, most autonomous jobs, where the model has to stay coherent across many connected steps over hours or days with dense source material," the spokesperson said, advising customers to "run both on a representative workload, one bounded task and one long-horizon job." That framing — bounded tasks versus long-horizon autonomy — may become the defining axis of model differentiation in 2026, as benchmarks saturate and the hardest remaining problems involve sustained, multi-day agentic work rather than discrete puzzles.

Why token efficiency is becoming the real battleground for enterprise AI spending

Threaded through the launch is a theme Anthropic clearly wants buyers to absorb: Opus 5 doesn't just score well, it scores well per dollar. The model ships with an adjustable "effort" setting that lets customers trade intelligence for speed and token savings, and Anthropic's charts emphasize performance at a given cost rather than peak performance alone.

Early customers echoed the point with unusual specificity. Harvey, the legal AI company, said Opus 5 achieved similar performance to Opus 4.8's maximum-reasoning mode "while generating 26% fewer tokens on average," according to Niko Grupen, its head of applied research. Richard Pham of Fundamental Research Lab said that on hard financial-modeling tasks, the model averaged nine percentage points higher accuracy "while using roughly one-third fewer turns and tool calls and 60% less time."

Wade Foster, chief executive of Zapier, said Opus 5 topped his company's AutomationBench leaderboard "without spending more tokens than prior Claude models," running a full churn-prevention workflow from start to finish. "Previous models didn't pass; Opus 5 hit 100%," he said. Scott Wu, chief executive of Cognition, the company behind the Devin coding agent, said that on FrontierCode 1.1, "Claude Opus 5 approaches Fable-level performance at half the cost," with particular strength in debugging and root-cause analysis.

The efficiency emphasis reflects commercial reality. Enterprise AI spending is no longer experimental, and inference costs — the price of actually running these models at scale — have become a board-level line item. 

Anthropic's business skews heavily toward API and enterprise usage; according to a February 2026 analysis by Contrary Research, Claude held roughly 40 percent of the enterprise large language model market by usage as of late 2025, and Claude Code alone had reached about $1 billion in annualized revenue. For a company whose customers pay by the token, a model that does more with fewer tokens is not a nice-to-have. It is the product.

Self-verifying AI agents and what they mean for the hidden costs of automation

Beyond the numbers, Anthropic is selling a behavioral story: that Opus 5 verifies its work and iterates until it succeeds. The company offered several examples from testing that read like small parables of machine stubbornness.

In one Frontier-Bench task, the model was asked to reconstruct a machine part as a 3D CAD model from a drawing it was intentionally given no way to view. Rather than fail, Anthropic says, Opus 5 wrote its own computer vision pipeline to extract the geometry from raw pixels — and did so repeatedly, while no competing model solved the task in five attempts. In another case, given a real bug in a popular open-source package manager, the model found the root cause and fixed an edge case the community's own patch had missed; a competing model patched only the symptom and declared victory. An engineer at a trading firm, the company says, used Opus 5 to build a market data feed for a new exchange in a single session and, finding no live feed to validate against, watched the model build its own test harness to check its parsing code.

Customers described similar behavior in the wild. Cristian Rivera, a staff software engineer at Stripe, said he gave the model "a chief-of-staff role over my dev environments" for a weekend: "it built its own monitor, drove each box, and pulled me in only for the judgment calls."

This is the capability enterprises actually care about, and it is worth dwelling on why. The gap between a model that produces plausible output and one that verifies its output is the gap between a demo and a deployable system. Most of the hidden cost of enterprise AI today is human review — engineers checking the machine's work. A model that reliably checks its own work compresses that cost, which is precisely why customers keep citing fewer turns, fewer passes, and less time rather than higher raw scores.

Inside Anthropic's safety strategy: capability gaps, classifiers, and model fallbacks

The launch also showcases Anthropic's increasingly intricate approach to safety — one that now involves deliberately not teaching its models certain skills. The company says its automated behavioral audit found Opus 5 to be its most aligned model to date, scoring 2.3 on overall misaligned behavior, lower than Opus 4.8, Sonnet 5, or Fable 5, with the lowest rates of deceptive behavior and the least susceptibility to being tricked into misuse.

On the capability side, Anthropic says it intentionally avoided training Opus 5 on cyber tasks, as it did with Opus 4.8. The model improved on them anyway — a side effect of general capability gains — and now nearly matches Mythos 5 at finding software vulnerabilities. But it remains far behind at exploiting them: on Anthropic's OSS-Fuzz evaluation, Opus 5 identified vulnerabilities at a 79.4 percent rate, close to Mythos 5's 80 percent, but succeeded at developing exploits in only 4 challenges versus Mythos 5's 13. That asymmetry — strong at defense-relevant discovery, weak at offense-relevant exploitation — appears to be by design, and the safeguards follow the same logic. Anthropic expects Opus 5's cyber classifiers to intervene about 85 percent less often than Fable 5's.

When a classifier does trigger, requests in Claude.ai, Claude Code, and Claude Cowork fall back to Opus 4.8 by default — raising an obvious question: if a request is too risky for one model, why is it acceptable for another? "The model it falls back to has lower capability levels making the risk of harmful use lower as well," the spokesperson said, adding that "there is a message that lets the user know when this occurs and is visible in the chat."

The logic is defensible, but it reveals how AI safety actually works in 2026: risk is not a property of the question alone, but of the question multiplied by the capability of the system answering it. On biology, the calculus runs the other way. Opus 5 is now Anthropic's most capable generally available model for scientific research — scoring 10.2 percentage points higher than Opus 4.8 on the company's internal chemistry benchmark — though the spokesperson acknowledged that "Mythos 5 remains the stronger model for long-horizon, open-ended work like autonomous drug design campaigns."

The business stakes behind the launch: a $380 billion valuation and massive compute bets

The launch lands at a moment of extraordinary commercial momentum — and extraordinary obligations — for Anthropic. Reuters reported in February that the company was valued at roughly $380 billion in its latest funding round, following a period in which, per Contrary Research's analysis, its annualized revenue climbed from about $1 billion at the end of 2024 to a projected $9 billion by the end of 2025, with internal targets reportedly reaching $20 to $26 billion for 2026. Those targets are underwritten by enormous infrastructure commitments, including a reported $30 billion Azure compute deal alongside arrangements with Google Cloud and Nvidia — spending that only pencils out if enterprises keep expanding usage.

That is the context in which Opus 5's pricing strategy makes sense. Holding the price at Opus 4.8 levels while roughly doubling performance on key agentic benchmarks is effectively a steep price cut per unit of capability, designed to widen the funnel of workloads that are economical to automate. Every task that was marginal at Opus 4.8's cost-per-success becomes viable at Opus 5's — and every viable task is recurring token revenue.

The regulatory backdrop has grown more complex as well. A U.S. judge gave final approval this week to Anthropic's $1.5 billion copyright settlement with book authors, Reuters reported, closing a chapter of litigation over the company's early training data. And in June, Reuters, citing Axios, reported that the U.S. government had moved to block foreign access to Anthropic's most advanced models — a reminder that frontier AI is now entangled with export policy in ways that shape which customers can buy what.

Also shipping Friday: a Fast mode running at roughly 2.5 times default speed at twice the base price, automatic fallback routing on the API, and mid-conversation tool changes that no longer invalidate the prompt cache — a small feature that agent developers may appreciate more than any benchmark. Consistent with prior Opus models, Opus 5 carries no data retention requirements for general access, a point the spokesperson flagged unprompted for customers with "a hard zero data retention requirement." Developers can access the model as claude-opus-5 on the Claude API starting today.

Two questions will determine whether the bet pays off: whether Opus 5's efficiency claims survive contact with production workloads at scale, and whether enterprises embrace a world where safety classifiers, not users, sometimes decide which model answers. But the deeper message of Friday's launch is that the AI industry's center of gravity has moved. For three years, the labs competed on what their best model could do on its best day. With Opus 5, Anthropic is competing on something less glamorous and far more lucrative: what a very good model can do every day, for half the price. In a market where the frontier keeps moving, Anthropic is wagering that the real fortune lies just behind it.

• Added Claude Opus 5 (claude-opus-5), now the default Opus model — 1M context, fast mode at $10/$50 per Mtok
• Added sandbox.network.strictAllowlist setting to deny non-allowlisted hosts for sandboxed commands without prompting
• Added DirectoryAdded hook that fires after /add-dir or the SDK register_repo_root control request registers a new working directory mid-session
• Added mcp_server_errors to the headless stream-json init event, listing --mcp-config entries skipped by config validation; terminal runs print a startup warning
• Added the workflowSizeGuideline settings key so the advisory Dynamic workflow size guideline can be set from any settings file; the /config row is hidden while one does
• Added nested subagent forwarding in stream-json: subagents spawned at depth-2+ now appear when --forward-subagent-text is set, keyed by their spawning Agent tool_use id
• Fixed claude -p text output dropping the answer already produced when a turn dies on a mid-stream API error
• Added HTTP status and error text to claude mcp list and /mcp when a server fails to connect, and a warning for MCP config values with hidden leading or trailing whitespace
• Fixed a permission you approved while a self-hosted runner was restarting being dropped when the session resumed, so the approved action now runs
• Fixed the Fable model row showing "Requires usage credits" for plans that include it, when a stale cache had baked the label in
• Fixed a SIGTERM arriving while a self-hosted runner was starting up leaving a stale active row until the lease expired; it now deregisters cleanly
• Added structured failure categories to self-hosted runner spawn and session failures, so hook errors, runner crashes and config errors can be told apart
• Fixed the /model picker showing the merged Opus row as plain "Opus" instead of "Opus (1M context)"
• Fixed copy-on-select inside GNU screen printing base64 into the terminal instead of copying the selection
• Fixed Remote Control clients keeping a stale fast-mode status after a model switch, reconnect, or failed org check
• Fixed CLAUDE_CODE_GIT_BASH_PATH on Windows exiting or being used as bash when the path isn't a bash/sh binary; it's now ignored with a warning
• Fixed Vim mode: pressing ← on an empty prompt now returns to the agent view from NORMAL mode, not just INSERT
• Fixed screen-reader mode rewriting the entire input line on every keystroke instead of echoing only the typed character
• Improved the "Remote Control is only available via api.anthropic.com" error to name the specific setting that caused it
• Improved claude --teleport to show which repo your current checkout points at when it doesn't match the session's repo
• Changed dynamic workflows to default to a medium size guideline (aim for fewer than 15 agents); pick another size or unrestricted with Dynamic workflow size in /config
• Changed managed MCP allowlist/denylist ${VAR} entries to resolve from the startup environment and managed-settings env instead of settings-file env
• Changed the /model picker to highlight only the newest model's name, so the highlight marks the new release rather than an arbitrary subset of the list
• Added the current default workflow size to the running-workflow status line, with a pointer to /config for changing it
• Removed Opus 4.7 from fast mode; /fast now applies to Opus 5 and Opus 4.8
• Updated the claude-api skill to default to Claude Opus 5, with a migration path from Opus 4.8
• Subagents can now spawn nested subagents up to depth 3 by default (was 1); set CLAUDE_CODE_MAX_SUBAGENT_SPAWN_DEPTH=1 to disable nesting

Read more: What is preventing employees from changing careers, asks Indeed?

The post AI for SMEs: closing the readiness-reality gap between ambition, execution and measurable value appeared first on SME Cybersecurity News.

Der Beitrag Apple TV: Neue Dating-Show und finale Staffel von „The Morning Show“ angekündigt erschien zuerst auf appgefahren.de.

The post Reddit Rethinks Google AI Partnership as AI Search Reshapes Publishing appeared first on TechRepublic.

The Freiburg, Germany-based AI lab says FLUX 3 is jointly trained across those modalities rather than assembling separate image, video and audio models behind a common interface.

That distinction is central to the company's pitch: BFL wants enterprises to think about creative generation, simulation, computer use and robotics as connected applications of a single capability it calls visual intelligence — models, in the company's words, "that can perceive, predict, and act across physical and digital environments." This release marks BFL's first public video generation model.

FLUX 3 will be offered through four product lines: FLUX 3 Video, FLUX 3 Image, FLUX 3 Action and the upcoming, open source FLUX 3 Dev. FLUX 3 Video, with optional native audio generation, and FLUX 3 Action are entering a gated "Early Access" program now, to which anyone can apply, but which BFL must approve.

There is presently no public access through BFL's application programming interface (API) or those of partners yet, but the company says FLUX 3 Image will roll out in the coming weeks, followed by general availability. The limited initial availability rollout echoes the release strategies of new models from other frontier labs in the U.S. lately, including Anthropic and OpenAI, though those were ostensibly for security concerns and due to government request.

What the company has not announced is pricing, production service-level commitments, evaluation methodology, sample sizes, rater counts or any image-model benchmarks at all. Enterprise buyers therefore cannot yet calculate total cost of ownership or independently reproduce the video comparisons.

Another big notable omission: FLUX 3 is not launching with downloadable weights at this time, nor an open source license. BFL says faster and open-weight versions will arrive later this year, and its technical blog names FLUX 3 Dev as "open-weight access to a multimodal backbone, for content creation (video, audio and image) and action prediction" — a considerably broader commitment than any previous FLUX Dev release, all of which covered images only.

But it arrives last in the sequence. Developers accustomed to receiving a locally deployable FLUX variant alongside — or soon after — a major model announcement will have to wait. That delay does not negate the company's commitment, but it is disappointing given the role open weights have played in FLUX's adoption thus far.

Flux 3 is rated higher than the competition, but missing pricing and benchmarking details may prevent rapid enterprise adoption

BFL has published several benchmark comparisons, but they're qualified as preliminary — with full benchmark results and methodology to be published later during broader general availability.

In early head-to-head preference testing on 10-second, 720p text-to-video clips with audio, the company says FLUX 3 was preferred over Luma Ray 3.2 in 93% of comparisons, Runway Gen-4.5 in 77%, Grok Imagine Video in 69%, Kling v3 Pro in 60%, Happy Horse v1 in 59%, Happy Horse 1.1 in 57%, and both Seedance 2.0 and Google's Gemini Omni Flash in 52%.

One caveat travels with every one of those figures, and it comes from BFL itself. The chart carrying the results is labeled a "preliminary evaluation of an early FLUX 3 candidate" — meaning the numbers describe a pre-release checkpoint rather than the model now entering early access. That cuts both ways: the shipping model may perform better, but nothing published today measures what customers will actually call.

Luma Ray 3.2 and Runway Gen-4.5, where FLUX 3 posted 93% and 77%, are the softest comparisons on the list — established products, but not the models currently setting the pace in independent video rankings. Those are real wins, and they are the ones least likely to change an enterprise shortlist.

Seedance 2.0, at 52%, is a statistical coin flip against a model most Western enterprises cannot currently procure. ByteDance indefinitely postponed Seedance 2.0's international rollout after Netflix, Warner Bros., Disney, Paramount and Sony sent legal threats over alleged systematic copyright infringement, and that suspension remains in place. Tying a frozen product is neither a strong claim nor a damaging one.

Gemini Omni Flash, also at 52%, matters much more. Omni is the closest large-platform analogue to what FLUX 3 is attempting — multimodal input, video and audio-aware creation, conversational editing — and by BFL's own measurement, the two are indistinguishable on 10-second text-to-video quality.

Google's advantage in that matchup is that Omni is generally available via Google's Gemini API for $0.10 per second of generated 720p video, or a 10-second clip for around.

One regional wrinkle matters for a German company's home market. Editing uploaded video is unavailable to Omni Flash users in the European Economic Area, Switzerland and the United Kingdom, though editing video the model itself generated is permitted. A European enterprise that wants to run its existing footage through a generative editing pass cannot currently do so on Omni Flash.

Here's a rough guide for enterprises considering which video models to rely upon:

One architecture for media generation and physical action

FLUX 3 builds on Self-Flow, BFL's method for aligning multimodal understanding and generation within one architecture, publicized back in March 2026.

The company says it significantly scaled up compute and data to train across video, images and audio simultaneously, and that testing showed video generation and action prediction do not require separate foundations — the same architecture could be extended to action prediction without sacrificing what it learned from video.

"We place vision at the center of our approach because it is the most signal-rich medium of the physical world. Images convey structure, images and video teach spatial relationships, video teaches dynamics, and actions reveal causal relationships. But vision alone is not the complete picture," said Robin Rombach, co-founder and CEO of BFL, in a pre-release statement provided to VentureBeat. "True intelligence means perceiving the world: predicting how it will change, taking action, and learning from the results. Joint training within one unified architecture is what will get us there, because each training modality strengthens the others. Audio conveys timing, prosody, and physical events that elude vision. Language conveys goals, abstractions, and instructions that pixels cannot easily express."

He put the case more bluntly elsewhere in the announcement: "You can't cheat reality. A model that only learns images can only generate images. But the world is not made of still frames. It moves, sounds, changes, and responds."

BFL says FLUX 3 targets creative tooling, media, design, e-commerce and physical AI, supporting video generation with synchronized audio, precise image editing, product and material consistency across motion, multilingual generation and robotic action prediction. It is already being tested by Canva, Burda, Magnific (formerly Freepik), Krea and Picsart.

For creative software companies, the appeal is consolidation. A single foundation could potentially support storyboarding, image editing, product rendering, video variation and localization without repeatedly translating assets and instructions between disconnected models.

For robotics teams, the potential value is data efficiency. Models that already encode motion, object behavior and physical change may need less task-specific robot training than systems starting from raw demonstrations.

What FLUX 3 Video can actually do

The video tier is the most concretely specified part of the launch, and it settles a question that had been circulating as rumor: FLUX 3 generates clips of up to 20 seconds with audio in a single generation.

Every video output comes with native audio. For comparison, HappyHorse 1.0 tops out at 15 seconds of 1080p with synchronized audio — though BFL has not stated what resolution its 20-second clips run at, and its published evaluations were conducted at 720p. Still, a 20-second long clip from a single prompt is among the longest yet achieved, matching OpenAI's discontinued Sora model.

The capability list BFL published covers:

• Text-to-video generation.

• Image-to-video generation, either animating from a starting frame or using images as visual references.

• Video-to-video generation from a reference clip, carrying elements such as a specific character into a new scene or context.

• Generative video-audio continuation from existing video and audio input.

• Keyframe-to-video generation for controlled transitions between defined moments.

• Multilingual dialogue.

• A broad range of visual styles and aspect ratios, from candid camcorder footage to animation and cinematics.

• Typography generation and animated design.

• Agentic chaining of individual clips into longer, multi-shot sequences.

That last item is the one enterprise video teams should look at hardest. BFL claims the capabilities combine to produce sequences lasting several minutes, with visual references keeping characters consistent across scenes. If that holds up under production conditions, it addresses the constraint that has kept generative video out of most commercial pipelines: not clip quality, but continuity across shots.

It is also the capability where competition is most direct. HappyHorse 1.1's headline upgrade is R2V, or Reference-to-Video, which accepts multiple character reference images to hold identity stable across generated footage — the same problem, approached at the input layer rather than through agentic clip chaining. Alibaba also claims zero-drift lip sync and has specifically targeted the artifacts that mark commercial AI video as synthetic, including facial oiliness and over-sharpening. Character consistency is where this category is being contested, and both companies know it.

BFL says FLUX 3 Video is already particularly strong at human facial expressions, associating sounds with physical events, and multilingual output. On the image side, the company says preliminary evaluations conducted during midtraining show significant improvement over earlier FLUX versions in complex prompt handling and text generation, including high-accuracy text in multiple languages. It published no image benchmarks or win rates.

FLUX-mimic tests whether video models can become robot models

BFL is applying its unified-architecture thesis through FLUX-mimic, a video-action model built on FLUX 3 and developed with Swiss firm Mimic Robotics, one of the first partners to receive early access.

The technical blog describes two distinct routes to action prediction: integrating native action prediction directly into FLUX 3, scaling up the initial Self-Flow work; and using the pretrained video backbone as a dynamics-aware foundation from which specialized action models can be finetuned with limited task-specific data. FLUX-mimic is the second route — the FLUX 3 backbone combined with mimic's robot-learning and production-deployment expertise in dexterous manipulation.

FLUX-mimic is designed for general-purpose robotic manipulation: helping robots understand a visual scene, predict the consequences of an action, and adapt to new tasks with far less task-specific data.

BFL and Mimic Robotics say that depending on task difficulty, the model can be finetuned for a specific manipulation task with as little as 30 minutes of robot data, where prior approaches have required 30 or more hours.

"The hardest part of robotics is data," said Elvis Nava, CTO of Mimic Robotics, in a statement provided to VentureBeat. "Every new task normally means hours of a robot repeating itself. Because FLUX-mimic is built on top of frontier video models that already understand how the physical world behaves, it picks up a new task in minutes, not days. This way, we can leapfrog the current state of the art in robot learning."

BFL argues that a model trained only on images cannot understand a world that "moves, sounds, changes, and responds," and that physical understanding is what produces convincing generated footage. Google makes a nearly identical claim for Gemini Omni.

Its developer documentation cites "world knowledge" that combines "an understanding of physics" with Gemini's grasp of history, science and cultural context. Its marketing is blunter still: "Most AI models just predict the next pixel to build a narrative or an image. Gemini Omni is different," the company posted in June, crediting the model with "an intuitive understanding of forces like gravity, kinetic energy, and fluid dynamics for more realistic movements that follow real-world logic."

The practical consequence for enterprise buyers is that world-model language is not a differentiator. Two of the three leading video systems now market physical understanding as their central advantage, and neither has published a benchmark that measures it.

There is no standard test for whether generated water behaves like water, whether a dropped object falls at a plausible rate, or whether a sound arrives when the impact does. Human preference ratings capture some of it indirectly. Nothing else on offer captures it at all.

Open weights helped make FLUX an industry standard

BFL officially launched in summer 2024 and gained a name for itself in the AI industry in the intervening two years for its commitment to open sourcing high-quality AI image models beloved by developers, creatives, and enterprises.

The company's founders, including Rombach, Andreas Blattmann and Patrick Esser, previously helped create VQGAN, latent diffusion and Stable Diffusion, the latter the open source technology that kicked off broad AI generation capabilities for the masses and currently used by many AI image generators and companies.

That reach translated into commercial distribution. FLUX models now power generative features inside Adobe Photoshop, Picsart and Nous Research's Hermes Agent, among other platforms, and the company cites film director Martin Scorsese among professional users.

Wired magazine described Black Forest Labs as a relatively small company that nevertheless became a leading competitor to Silicon Valley's largest AI labs, with FLUX models ranking near the top of image benchmarks and becoming some of the most downloaded text-to-image models on AI code sharing community Hugging Face. The company says it now runs a 100-person team across Freiburg and San Francisco.

FLUX.1 Dev, FLUX.1 Kontext Dev, FLUX.1 Fill Dev and related control models, released shortly after the firm's launch, gave researchers and creative-tool developers access to downloadable checkpoints, local inference and integrations with frameworks including Hugging Face Diffusers and ComfyUI. FLUX.1 Kontext Dev, for example, was released as an open-weight model for research and noncommercial use, with generated outputs permitted for commercial purposes under the applicable license.

The company continued that pattern with FLUX.2 Dev in late 2025, a 32-billion-parameter open-weight model combining generation and multi-reference editing. Black Forest Labs called it the strongest open-weight image generation and editing model available at launch and released weights, reference inference code and optimized implementations for consumer Nvidia GPUs.

FLUX 3 Dev raises the stakes on that evaluation. Previous Dev releases were image models. This one is described as a multimodal backbone spanning video, audio, image and action prediction — meaning a single license will govern whether a company can locally deploy a model that touches both content production and physical machinery. BFL hasn't yet shared information about its license, the parameter count, quantizations or hardware requirements.

The company frames open weights as an enterprise feature rather than a community gesture, arguing they enable secure, low-latency local deployment for applications like robotic control systems and let teams adapt FLUX 3 to their own data, products and workflows.

The financial backing behind FLUX 3 is worth noting alongside the technical claims. Black Forest Labs is valued at $3.25 billion and has raised more than $450 million from investors including a16z, AMP, Salesforce Ventures, Nvidia, General Catalyst, Adobe Ventures, Figma Ventures, Canva and Deutsche Telekom's T.Capital.

This wave of VentureBeat Pulse Research examines enterprise agent orchestration: which platforms enterprises run on, what drives the choice, what they optimize for, how they expect agent control to be structured, and — most revealingly — how orchestrated their deployed “agents” actually are and how tightly they control the cost of running them.

The central finding is a gap between orchestration ambition and orchestration reality. Enterprises are consolidating fast onto the major model platforms: Anthropic’s Claude is the primary platform for 40%, more than double any rival, followed by Microsoft (18%) and OpenAI (13%). The choice is driven by “model gravity” — native alignment with a state-of-the-art base model (21%) — and success is judged by reliable, multi-step execution (task completion reliability 32%, multi-step workflow management 28%). Yet asked to assess their portfolios honestly, 71% say a quarter or fewer of their deployed “agents” are true multi-step orchestrated workflows rather than single-prompt chatbot wrappers, and only 10% have crossed the halfway mark. The orchestration layer is being built well ahead of the orchestrated portfolio it is meant to run.

That gap shapes the architecture enterprises are putting in place. By the end of 2026 a clear majority (51%) expect a hybrid control plane — provider-native plus external orchestration — and only 6% expect to hand control to a provider-managed service, because vendor lock-in (35%) is the risk they fear most if control lives inside a model provider. Investment follows the build-out: agent workflow tooling leads the spend (34%), with security and permissions enforcement (25%) behind. And fiscal control lags throughout — more than a quarter (27%) have no real-time way to stop a runaway agent before the bill arrives.

Methodology

VentureBeat fielded this survey as part of its ongoing Pulse Research series, this instrument focused on enterprise agent orchestration. Responses are filtered to organizations with 100 or more employees (n=101), drawn from a single June 2026 wave; because this is one wave rather than a pooled multi-month sample, the report reads cross-sectionally and does not infer month-over-month trends.

By organization size the sample is spread evenly across the enterprise bands: 100–499 employees, 2,500–9,999, and 50,000+ (21% each), with 10,000–49,999 and 500–2,499 (19% each). By role it is senior and buyer-credible: product and program managers (15%), CIO/CTO/CISO (13%), consultants and advisors (13%), and a spread of data, AI, and engineering directors and VPs, with an “Other” function at 18%. On purchasing, 81% are recommenders, influencers, or final decision-makers for AI solutions (66% recommender/influencer, 15% final decision-maker). Technology/Software is the largest industry at 44%, followed by Financial Services (17%) and Healthcare/Life Sciences (8%).

At 101 respondents the sample is robust enough to read directionally with reasonable confidence, though it remains self-selected and is not a probability sample.

Finding 1: Orchestration runs on model-provider platforms

Anthropic’s Claude leads; open frameworks are marginal

We asked which agent orchestration platform enterprises primarily use today. The answer concentrates on the major model providers — and on one in particular.

A note on reading these shares. As described in the methodology section, the respondents are self-selected, and this question asked them for a single primary platform — so the figures measure which platform leads each enterprise's deployment, within a self-selected audience of AI-active technical decision-makers. A sample built this way can diverge substantially from spend-weighted market measures, and each VB Pulse survey draws its own sample with its own company-size mix, so vendor figures should not be compared across our surveys either. Read these shares as a portrait of where this cohort has placed its primary orchestration bet today, rather than as market share.

The model platforms dominate. Anthropic, Microsoft, OpenAI, Google, and Amazon together account for roughly 80% of deployments (81 of 101), while the open frameworks (LangChain/LangGraph) and custom in-house builds that anchor engineering discussion sit in single digits. Anthropic’s lead — 40%, more than double the next platform — mirrors the “model gravity” selection logic in Finding 2: enterprises are choosing the orchestration layer that comes with the model they want to build on. As with the security vendors in the prior agent-security wave, the tools that define the category in technical circles are not yet where enterprise deployment concentrates. A small 3% are not orchestrating at all.

Respondents rate the platforms they run at 3.94 out of 5 overall (109 answered), with “value for money” specifically at 3.94 and “ease of implementation” the weakest score, at 3.85 — placing orchestration near the bottom of our five-tracker satisfaction range, ahead of only evaluation tooling. A rating just under 4 out of 5, from users of whom 96% plan to change their orchestration approach within the year, reads as provisional acceptance: the platforms work well enough to run today, and not well enough to stop the search for something better. The ratings sit alongside near-universal intent to change; this is a layer enterprises tolerate more than they love.

Finding 2: Model gravity drives platform selection

The base model, not the tooling, decides the platform

We asked what most influenced the orchestration platform choice. The single largest factor is the pull of the underlying model — though flexibility and ease of development follow close behind.

Model gravity leading is the selection-side explanation for Anthropic’s platform lead: enterprises pick the orchestration environment closest to the frontier model they have standardized on. But the next tier complicates the picture — flexibility across models and tools (17%) and ease of development (17%) say enterprises also want to avoid being trapped by that choice, foreshadowing the lock-in fear in Finding 6. Security and permissions (14%) and total cost of ownership (11%) round out a pragmatic buying logic. Performance (latency/memory) sits last at 4%, a reminder that at this stage of adoption the binding constraints are model fit and optionality, not raw speed.

Finding 3: The job is reliable multi-step execution

Enterprises just orchestration by whether it completes the work

We asked what enterprises optimize for — their primary success metric for orchestration. Reliability and multi-step workflow management dominate; developer- and user-facing metrics trail.

Task completion reliability (32%) and multi-step workflow management (28%) together account for 59% of responses (60 of 101): orchestration succeeds, in the enterprise view, when it reliably carries a task through multiple steps to completion. Developer productivity (17%) matters but is secondary — the inverse of its prominence in framework discussion — and end-user experience (9%) is a minor concern, consistent with orchestration being an internal execution problem rather than a UX one. This reliability-first standard is exactly what makes the Chatbot Trap finding so pointed: enterprises define success as dependable multi-step execution, yet most of their deployed “agents” do not yet do multi-step work at all.

The trap is not evenly distributed. Splitting the sample by organization size, 77% of smaller enterprises say a quarter or fewer of their agents do true multi-step work, against 62% of larger ones. Larger enterprises are meaningfully further into genuine multi-step deployment; the chatbot trap is, directionally, a mid-market condition.

Finding 4: Consolidate, productionize, and build in-house

Three strategic moves are nearly tied for the year ahead

We asked what major change enterprises anticipate in their orchestration strategy over the next 12 months. Three moves cluster at the top, almost evenly split.

The top three — building in-house control (25%), standardizing on one framework (24%), and moving agents from sandbox to production (23%) — are statistically indistinguishable and tell a single story: enterprises are moving from experimentation to operational consolidation. They want fewer frameworks, more production exposure, and more ownership of the control layer; only 4% expect no change. The appetite for custom in-house control planes is notable alongside the platform concentration in Finding 1 — enterprises are standardizing on model-provider platforms while simultaneously planning to wrap them in control logic they own, the hybrid posture that Finding 6 makes explicit.

Finding 5: Nearly seven in 10 plan to switch — and the biggest group of movers has no shortlist 

The strategic change enterprises anticipate (previous finding) comes with vendor motion attached. Asked whether they plan to adopt a new, additional, or replacement agent orchestration platform in the next twelve months, more respondents are moving here than in any other layer we track.

Asked which platforms they are considering, the most common answer among those in motion is none yet: 29% of all respondents are evaluating without a shortlist, the largest single response after "not considering a change." Among named candidates, OpenAI leads at 16%, followed by LangChain/LangGraph at 12% and Anthropic at 7% — and notably, the independent frameworks draw roughly double their current usage footprint in forward consideration, the same pattern our security tracker found for specialist vendors. Read with this report's concentration and lock-in findings, the picture completes itself: the major model-platform providers hold roughly four-fifths of today's primary usage, vendor lock-in has become the leading fear, 96% anticipate a strategic change — and now the purchase intent to act on all of it, with the largest bloc of buyers still undecided. The most concentrated layer of the agentic stack is also, as of June, the least settled.

Finding 6: Investment flows to workflow tooling

Tooling and permissions lead the spend; monitoring trails

We asked which orchestration-related investment will grow most next year. Agent workflow tooling leads, with security and permissions enforcement behind.

Workflow tooling leading (34%) is the budget-side expression of the reliability-and-multi-step priority in Finding 3: the money is going to the machinery that strings steps together dependably. Security and permissions enforcement (25%) and scaling infrastructure (20%) follow — the investments required to take agents from sandbox into production, the strategic move in Finding 4. Monitoring and debugging draws a smaller 11%, with another 11% reporting flat budgets. The weight on tooling, permissions, and scaling over pure observability signals that enterprises are spending to build and harden orchestration, not merely to watch it run.

Finding 7: The control plane will be hybrid — and lock-in is why

Enterprises expect to split control between providers and their own layer

We asked where enterprises expect the primary control plane for agents to live by the end of 2026, and what worries them most if that control sits inside a model-provider platform. A clear majority expect a hybrid model — and vendor lock-in is the reason.

Hybrid control is the dominant expectation by a wide margin (51%), and only 6% expect to hand control to a provider-managed service outright. Read together, the hybrid, custom, and externally-abstracted options — every architecture that keeps control at least partly outside the provider — sum to 88% (89 of 101). The reason surfaces directly when we asked about the risk of provider-resident control: vendor lock-in leads at 35% (35 of 101), ahead of security and permissioning limitations (28%) and inflexibility across models and tools (21%). The pattern echoes the prior wave’s “don’t trust the model to police itself” posture — here, enterprises will build on a provider’s platform but decline to be governed entirely by it. The hybrid control plane is the architectural hedge against the lock-in they most fear.

The June figure asserting a preference for a hybrid control plane marks movement from earlier. In the April–May survey (n=145), only 34% expected a hybrid control plane, and a greater number (12%) expected to hand control fully to a provider-managed service. These two snapshots don’t yet measure a confirmed longitudinal trend — but the direction of the conversation is unambiguous: toward keeping control.

Lock-in is also a new arrival as a top concern. In the April–May wave, the leading concern was security and permissioning limitations (32%), with lock-in second at 24%; by June the two had traded places. The worry about provider platforms appears to be maturing from whether they can be secured to whether they can be replaced.

Finding 8: The chatbot trap — most “agents” aren’t agents yet

Enterprises admit most deployments are still chatbot wrappers

We asked enterprises to assess their portfolios honestly: what share of their deployed “agents” are true multi-step orchestrated workflows versus simple single-prompt chatbot wrappers. The answer is the defining finding of this wave.

This is the gap at the center of the report. Combining the bottom two bands, 71% of enterprises (72 of 101) say a quarter or fewer of their deployed “agents” are genuinely orchestrated — and just 10% (10 of 101) have crossed the halfway mark. The ambition documented in the earlier findings — model-provider platforms, reliability-first success metrics, production rollouts, a deliberate control architecture — runs well ahead of the deployed reality, which remains overwhelmingly single-prompt assistants dressed as agents. This is less a contradiction than a roadmap: the platforms, budgets, and strategies are being put in place precisely because the orchestrated portfolio is still so thin. The open question for later waves is how fast the reality closes on the ambition.

Finding 9: Fiscal control is still reactive

Only a minority can stop a runaway agent before the bill arrives

Finally, we asked how enterprises enforce fiscal control over agent token consumption — the risk that an autonomous loop exhausts a budget before anyone intervenes. Most rely on native caps or after-the-fact monitoring; real-time programmatic control is the exception.

More than a quarter of enterprises (27%) admit they have no real-time, programmatic way to stop an agent before a budget-breaking bill arrives — they learn of it from the logs afterward. Another 32% lean entirely on the native caps and throttles built into their primary platform, a control only as good as the provider’s tooling and one that ties back to the lock-in concern of Finding 6. The enterprises building custom gateways (23%) or exploiting cross-model routing to arbitrage cost (19%) are the ones treating token burn as an engineering problem to be controlled deterministically. As with orchestration maturity, fiscal control is an area where the operational reality lags the ambition: agents are moving toward production faster than the cost-control plane around them is being built.

It’s worth noting, a split appears according to company size: roughly one in three enterprises under 2,500 employees (34%) exercises only reactive control of agent spend, against 20% of larger enterprises — directional figures, but consistent with the chatbot-trap split. The mid-market is running the least mature agents on the least instrumented budgets.

The bottom line: The layer is real; most of the agents aren't yet

Organizations with 100 or more employees describe an orchestration strategy that is consolidating quickly and maturing slowly. They are standardizing — for now — on model-provider platforms, which collectively hold roughly four-fifths of primary usage, chosen for the gravity of the underlying model, and they judge success by reliable multi-step execution. Investment is flowing to workflow tooling and permissions, the strategy is to consolidate frameworks and push agents into production, and the control plane they expect is deliberately hybrid, because vendor lock-in is the risk they fear most. But the standardization is provisional: 68% plan to adopt a new, additional, or replacement orchestration platform within twelve months — the highest switching intent of any layer we track — and the largest group of those movers has not yet shortlisted a candidate. Today's concentration describes where enterprises are, and visibly does not describe where they intend to stay.

But the honest self-assessment punctures the ambition. Seventy-one percent say a quarter or fewer of their deployed "agents" are truly orchestrated, only 10% are past the halfway mark, and more than a quarter cannot stop a runaway agent in real time. The orchestration layer — the platforms, the budgets, the control architecture — is being built ahead of the orchestrated portfolio it is meant to run. At 101 respondents in a single June wave this reads as a clear directional signal rather than a precise measurement: enterprises have decided how they want to orchestrate agents well before most of their agents are doing anything an orchestration layer is for. The questions for subsequent waves are whether the deployed reality closes the gap on the ambition — and, with nearly seven in ten buyers in motion and most of them undecided, which platforms the settled stack finally lands on.

Based on survey responses from 101 qualified enterprise respondents (100+ employees), drawn from a single June 2026 wave. Because this is one wave rather than a pooled multi-month sample, results read directionally rather than as a confirmed trend. Respondents include product and program managers, CIOs, CTOs and CISOs, consultants and advisors, and directors and VPs of data, AI, and engineering, across Technology/Software, Financial Services, Healthcare, and other sectors.

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.

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: