In what they call the first-ever real-world agent-to-agent exploitation method, Pillar Security researchers say they discovered an exploit in the repository behind Google's Agent Development Kit for Python that could allow attackers to compromise supply chains. In other words, now we know that one AI agent can be used to control and compromise another one that has more privileges. The security snafu existed in google/adk-python, an open source Python toolkit with more than 90 million downloads used to build and deploy AI agents. Google has since fixed the underlying issue in the repository but deemed the exploit non-rewardable because it involved social engineering. Even so, it illustrates the risks of using AI agents in CI/CD workflows for triage, pull request (PR) reviews, and discussions. It also shows how one AI agent could attack another in a production environment, according to Pillar’s Dan Lisichkin, who found and reported the vulnerability. “Our world is changing quickly, and new attack surfaces are not yet reflected in threat models because these attacks never could exist in the first place in the ‘pre-agent’ world,” Lisichkin said in a technical write-up published on Monday. He will also discuss the findings during a poster talk at DEF CON's AI Village on Friday, August 7 at 1600 PDT. “CISOs and security practitioners should start considering these scenarios, threat-modeling them, and calculating worst-case implications and blast radius,” Lisichkin wrote. The issue stems from the way that the repo ran two classes of automated AI agents with different privilege levels that unintentionally share a trust boundary. One is a low-privilege, public-facing AI agent activated whenever a user opens a pull request (PR) or issue, and a second is a high-privilege, maintainer-only agent. Pillar’s team found that the low-privilege, public-facing agent could be manipulated via prompt injection into triggering a maintainer-only agent that can execute malicious actions. “Because workflows that explain how these agents work behind the scenes are also public, any person could have connected the dots that one agent should be able - at least theoretically - to 'call' the other,” Lisichkin told The Register. "When it comes to building the attack, you just need to know English to build the prompt injection (or just ask an AI to do it for you)." There is one caveat: an attacker would first likely need to make legitimate contributions to the repository to build trust among the maintainers before moving on to prompt injection. But assuming someone was willing to put in the time, here’s how the attack would play out. First, an external user - this would be the attacker - creates a new PR. Lisichkin calls this PR A, and it combines a real fix with malicious code, such as a modified package.json or malicious dependency. Then, a public-facing agent tied to a high-privilege collaborator personal access token (PAT) reads the attacker’s PR text and marks the PR for review. This level of trust - the collaborator PAT - allows the attacker-generated text to trigger a gated workflow. Once the PR A triage happens, the attacker opens a second PR - PR B - with the prompt injection, and the triage agent emits the trusted @gemini-cli handoff. This triggers the privileged-agent workflow and executes the malicious action. “Strung together, they manufacture a complete, believable ‘a human asked for a review, gemini ran it, gemini approved’ trail on the poisoned PR, none of which ever happened,” Lisichkin wrote. Google did not respond to The Register’s inquiries, but Lisichkin confirmed that the underlying issue was fixed. Still, his findings, Google said, “did not meet the bar” for a bug-bounty payout. “This report demonstrates exfiltration of a GitHub token with a 'pull-requests: write' permission, which enables tampering with a PR but still requires a maintainer to take an action to merge the malicious PR as PRs are not automatically merged after a bot review,” Google explained. “We don't reward vulnerability reports that require social engineering to enable a supply chain security compromise,” the rationale continued. “Nonetheless, we have taken an action to harden the repository so we will be recognizing this report with credit.” Lisichkin told us the research shows agent isolation is not enough. "Agents should have their own identity, which mandates what resources they are allowed to access and in what they are allowed to interact with these resources," he said. "In this case, if Google had just given a bot identity to the initial triaging agent, most of the attack could have been prevented. Security teams need to start modeling agent identity and agent resource access within their threat models."®

Now AI is making fake vulnerabilities and polluting the ecosystem. A batch of critical- and high-rated SQLite CVEs that appeared in the NVD with CISA-supplied enrichment last week turned out to be technically bogus, according to security researchers, and their path into widely used databases exposes weaknesses in the CVE pipeline. Software supply chain security outfit JFrog reported last week that six supposed SQLite vulnerabilities published in a larger batch by a new, obscure GitHub repository were all complete garbage. Running the advisories through an AI checker suggested they were likely AI generated, JFrog said, and, upon testing, it found that none of the six SQLite reports, which carried CVSS scores ranging from 9.8 to 7.5, described a reproducible vulnerability. One, an alleged use-after-free vulnerability in the open source database that Red Hat initially assigned a maximum 10.0 CVSS score to before lowering it, relied on a function that didn't exist in the affected SQLite version. Another UAF vulnerability with a 9.1 CVSS score cited source lines that weren't even related to the supposed flaw. When JFrog tested the accompanying proof-of-concept, it executed a valid query with no memory leaks or errors. The other four SQLite CVEs from the repo that JFrog tested were similarly fake. The other 49 CVEs in the questionable GitHub repo claimed to be security vulnerabilities in the open-source RAW image processing library libraw and Arduino audio decoding library ESP32-audioI2S. While JFrog didn't test those as extensively, it said all are just as fake as the rest, aside from one which “contained a real bug wrapped in unverified CVE metadata.” A message posted to Openwall’s OSS-Security mailing list on Friday indicated that MITRE had rejected the whole repo’s worth of vaporous vulnerabilities, but the whole thing should serve as an important lesson, poster and Oracle Solaris engineer Alan Coopersmith pointed out. “MITRE and most other CNAs which assign CVEs for code they don't produce themselves operate on the honor system, and trust CVE requesters to have verified the information they provide,” Coopersmith noted in the OSS-Security post. “The CNA is often not in a position of being able to verify the report themselves.” As JFrog points out, the US National Institute of Standards and Technology (NIST), which manages the US National Vulnerability Database (NVD), used to provide a reliable backstop by manually reviewing and enriching CVE records after they entered the database. That process slowed dramatically in 2024 after a surge in vulnerability submissions, coupled with operational challenges, left the agency with a growing backlog of records it was unable to process. By late 2024, the backlog had grown to more than 17,000 unprocessed CVEs, despite NIST's plan to clear it by the end of fiscal year 2024 with contractor help. It continued to grow, reaching more than 27,000 by the end of 2025, according to a Department of Commerce Inspector General report published in May 2026. To make matters worse, the DoC IG concluded that NIST had been wasting money allocated to fixing the backlog due to a “lack of strategic planning and decisive action” that has led to the stack of unresolved issues continuing to grow. In other words, the pipeline has no mandatory checkpoint at which every claimed vulnerability must be independently reproduced. “Because no step in today's system actually requires a proof-of-concept or bug reproduction, a plausible-sounding fake advisory can slide right through the pipeline and end up in GitHub Security Advisories, downstream databases, and enterprise scanners,” JFrog said. “This incident demonstrates a systemic issue with automated vulnerability ingestion.” What that means for security professionals, aside from having to deal with polluted vulnerability databases, is that bad advisories could waste time better spent chasing real issues. Because reputable databases can ingest unverified records, JFrog recommended several checks before defenders act on a newly published CVE. First off, if the vendor hasn’t corroborated the issue (SQLite maintainers don’t list the fake CVEs, for instance) it’s probably not legitimate. A lack of commit hash or pull request in the reference fields of a repo is also indicative of AI slop, as is suspicious metadata (i.e., missing CPE product definitions). Lastly, if the code references don’t appear to match real functions or point to parts of the code that don’t involve the supposed issue, that’s a good sign it’s just an AI hallucination. JFrog reported its findings to the GitHub Security Advisory team, Red Hat, and NVD, all of whom the company told us have flagged or removed the CVEs. GitHub hasn't yet, JFrog told us. We reached out to GitHub to inquire why the repo is still up, but didn’t hear back. As for why someone might do this, JFrog speculates that it could be an attempt for someone to boost their research experience with fake reports, or to influence what automated CVE identification tools flag as actual vulnerabilities. In both cases, JFrog told us, that's just speculation. Either way, these 54 apparently bogus CVEs, JFrog security researcher Afek Berger said, are just one example of a problem they expect to see more often. "Generative AI has lowered the effort required to produce a plausible-looking advisory to close to zero, while the effort required to verify one, review the source code, build the affected version, reproduce the PoC, is unchanged," Berger told us in an email. "That asymmetry means that even well-resourced defenders and maintainers cannot manually validate every incoming report … this is a challenge the whole industry is facing in the AI era." ®

Conference-goers may want to think twice about connecting to public Wi-Fi after Microsoft disclosed that Russian foreign intelligence operatives (SVR) are compromising captive portal networks to deliver infostealers, keyloggers, and other malware. With the help of ReliaQuest's earlier work, Redmond fingered Storm-2945, a subdivision of the SVR's Midnight Blizzard (aka Nobellium), in an attack campaign targeting users of public Wi-Fi networks at places like hotels, conference centers, and other shared venues in the hospitality sector. Microsoft is still trying to determine how the hackers initially compromise captive-portal networks. The broader AI-assisted operation dates to February 2026, with traffic manipulation observed since early May. After gaining control of the network layer, Storm-2945 manipulates DNS and HTTP traffic to reroute users through attacker-controlled infrastructure, Microsoft said. The crew also abuses operating systems' connectivity checks to trigger malicious prompts and redirects. This gives the attackers an adversary-in-the-middle (AitM) position. Such prompts adopt ClickFix-style methods, which in some cases try to convince public Wi-Fi users to install malware under the guise of OS updates, driver repairs, and web verification failures. Users who follow through on the instructions provided in the prompts may then find their device infected with malware. Microsoft calls the campaign "CaptiveCrunch." One of the malware strains it delivers is CornFlake. Described as "a full-featured Windows RAT" written in Go, CornFlake is the SVR's go-to persistent implant in these hospitality network attacks. After presenting users with a "convincing" fake Windows update progress window, it provides attackers with a wealth of capabilities once installed. These include: Keylogging Clipboard monitoring Screenshot capture Audio surveillance Video surveillance Browser credential theft File exfiltration USB drive monitoring Security posture sweep Remote shell Microsoft also said that CornFlake exposes a localhost HTTP API server to transform the malware into a modular platform, delivering additional payloads such as ChocoShell, a PowerShell-based infostealer. ChocoShell is delivered and executed entirely in-memory, Microsoft said. SVR uses it primarily to suck up victims' browser session cookies, saved passwords, SSO tokens, and Wi-Fi credentials. Microsoft neatly summarized the two: "Where CornFlake provides the operator with a persistent, long-running foothold on the device, ChocoShell is designed to extract the most operationally valuable credentials, giving the operator access to victim cloud environments." The attacks primarily target Windows machines, but Microsoft has also seen indications of ClickFix prompts tailored to Android devices, encouraging users to download and install an APK file. In addition to the malware element, "a portion" of SVR's CaptiveCrunch activity is devoted to device code phishing. Users sent to attacker-controlled landing pages may be instructed to enter a device code on a legitimate Microsoft authentication page, unwittingly authorizing the attacker's session. Device code phishing exploits a legitimate OAuth flow, typically reserved for devices that struggle to open browsers, such as smart TVs. In such scenarios, attackers request an authentication code from Microsoft, which they then send to phishing targets. In the CaptiveCrunch campaign, this looks like a fake landing page, served to the user thanks to the AitM component of the attack. Targets are then asked to copy the code, which was originally given to the attacker, open a legitimate Microsoft authentication window, enter the code, and choose which account they wish to authenticate. Choosing the account completes the authentication flow, but in turn authenticates the attacker into the chosen account. This gives the attacker a valid OAuth token for the victim's Microsoft 365 account, potentially granting access to cloud data permitted by the token until it expires or is revoked. Device code phishing is not a new or unique attack, but can be an effective route to bypassing MFA, especially when an attacker already controls the flow of traffic after a captive portal compromise. "This activity is consistent with previously reported device code phishing operations conducted by Midnight Blizzard since August 2024," Microsoft said. "The observed technique does not appear fundamentally novel; however, integrating device code phishing into captive portal and traffic manipulation operations might increase the likelihood that users perceive the authentication request as legitimate." The main takeaway, in Microsoft's book, is to stop trusting public Wi-Fi so much. It did not discourage using hospitality networks' Wi-Fi services altogether, but said favoring personal hotspots and satellite internet connections over public networks is a safer bet. The majority of Redmond's advice could be brought under the user education umbrella: Don't trust public networks; teach users not to download updates over public networks or via prompts; educate users about what ClickFix attacks look like. That sort of stuff. But organizations have a role to play too. Among other technical implementations, passwordless authentication can thwart many phishing techniques, although device code phishing may bypass even passkeys. The best response would be for an employer to disable the device code authentication flow altogether, wherever possible, preventing staffers from surrendering their workplace cloud access to attackers. ®

Georgia and Michigan are the latest US states to report cyberattacks on water systems, as the FBI investigates incidents across at least seven states. Iran-backed hackers are the leading suspects, although the bureau has not publicly attributed the campaign. Officials in both states told journalists over the weekend that water facilities had detected activity consistent with the attacks on more than 30 Minnesota sites last week. Neither state reported operational disruption. Nine Michigan water systems reported hostile cyber activity to the state's Department of Environment, Great Lakes, and Energy. Department communications director Dale George said the state received "a small number" of reports consistent with the activity seen in Minnesota, but no public health consequences followed. "All systems continued to operate safely, issues were addressed by local operators, and there are no known impacts that posed a public health concern," said George. Georgia also confirmed to ABC News that it was affected, but said the damage was limited. Neither Georgia nor Michigan has published any form of public-facing notification about the cyberattacks. The three states are among at least seven affected by the intrusions, according to an FBI advisory posted last week. The bureau did not name a culprit or mention Iran. "Since 27 July 2026, Water and Wastewater Sector (WWS) utility companies in at least seven states have reported incidents to the FBI, and some of that activity degraded water operations," it stated in its advisory. The FBI said it had so far observed the activity only against Rockwell Automation/Allen-Bradley programmable logic controllers (PLCs), although it warned organizations deploying other manufacturers' devices to follow the same hardening advice. A broader CISA advisory, updated on July 22, warned that Schneider Electric, Siemens, and potentially other PLC brands were also being targeted by Iran-affiliated actors. Security researchers at Tenable were among the first to publicly suspect Iran's involvement, citing similarities with previous attacks by the IRGC-linked CyberAv3ngers group. Minnesota was the first state to confirm it was hit by the attacks, which took place over July 26-27. The state's IT department (MNIT), said more than 30 community water systems were targeted, but still has not officially attributed the attacks. According to WIRED, a restricted WaterISAC notice shared with water utilities said the Minnesota activity aligned with an earlier Iran-affiliated campaign. WaterISAC told WIRED that it had not assessed attribution "at any time" and publicly stated that it had not supplied the leaked document to the publication. President Trump also rejected the Iran link, offering no evidence for his alternative explanation. He told reporters following a cabinet meeting on Friday that "they blame it on Iran. I don't think so. I blame it on Minnesota because they're grossly incompetent." He added: "I think the governor is behind it. I don't think there was an Iranian cyberattack." Tim Walz, Minnesota's Democratic governor, suggested Iran was indeed behind the attacks, and highlighted Trump's funding cuts leaving sites such as water facilities more vulnerable to cyberattacks. "Trump knows exactly who is responsible for this attack, and knows that other states were hit too," he said. "This is what modern warfare looks like, and it further illustrates there's no plan to win a war in Iran. "DOGE took an axe to CISA and left the US exposed to cyberattacks. Thankfully, our experts in Minnesota were able to identify the vulnerability quickly and work with local communities to stop it." ®

The UK government's corporate finance adviser has admitted that an employee left an internal file containing the names and work email addresses of dozens of officials publicly accessible for around 40 hours. The breach, first reported by The Guardian, was disclosed in UK Government Investments' (UKGI) annual report, which says it occurred during the 2025-26 financial year after a member of staff "did not follow established information security policies." The exposed document contained "high-level management information" alongside the names and work email addresses of 51 government officials. UKGI, the Treasury-owned outfit that advises ministers on everything from corporate rescues to billion-dollar share sales, said it voluntarily reported the incident to the UK's Information Commissioner's Office even though it did not meet the threshold for mandatory notification. It also informed its Audit and Risk Committee and commissioned an external review of the breach. The report offers little else in the way of detail. UKGI doesn't say when the exposure occurred, where the file was hosted, whether anyone accessed or downloaded it, or which departments employed the affected officials. It also doesn't identify the external firm that reviewed the incident or disclose the recommendations it made. The review concluded that UKGI's response was appropriate and recommended further improvements to its security controls and incident preparedness. According to the report, "the overwhelming majority" of those recommendations have either already been implemented or are due to be introduced in the coming months. The mishap comes in a year when UKGI had its fingerprints on some of Whitehall's biggest commercial deals, from finally offloading the government's remaining NatWest shares to advising on small modular reactor financing and supporting the Eutelsat capital raise and Royal Mail takeover. The Register has asked UKGI for further details, including what information the file contained beyond names and email addresses, where it was publicly accessible, whether there is any evidence it was accessed while exposed, and what additional safeguards have since been introduced. Whether this was merely embarrassing or exposed officials to a meaningful risk depends on details UKGI has yet to disclose. An ICO spokesperson said: “We can confirm UK Government Investments Ltd reported an incident and we are assessing the information provided.” ®

A leading name in home and business physical security, Brinks Home, recently said it identified unauthorized access to a portion of its IT systems, an intrusion ShinyHunters claims it carried out to steal millions of records from the security provider's Salesforce instance. Brinks Home hasn’t named the intruder or identified the affected system, but said the responsible party has threatened to leak information it claims to have taken. “Brinks Home is working diligently to determine what information was involved and who may be affected,” the company statement said. “If the Company determines that personal information has been affected, it will notify those individuals as required and as appropriate.” An FAQ page for the incident said that Brinks Home products and services weren’t affected, as far as the company knows at this point, so alarms and other security tools should be working without issue. While Brinks may not have been very forthcoming with information, and lacks any sort of official way for media to communicate with it outside of sending a LinkedIn message it didn’t answer, the party that’s claimed responsibility has gone public with some details, and it’s none other than ShinyHunters with another claimed Salesforce breach. According to leak site monitoring outfit Ransomware.live, ShinyHunters claimed to have obtained more than 4.9 million Salesforce records from Brinks Home “containing some PII.” The group threatened this week to leak the data along with causing “several annoying digital problems” if Brinks Home didn’t reach out by Thursday, July 30, to negotiate a ransom payment. It’s not clear if Brinks Home has contacted ShinyHunters; Brinks Home didn’t respond to messages, and contacts The Register has for ShinyHunters appear to have changed, causing message and email rejections. ShinyHunters has been a prolific Salesforce intruder of late, with the group claiming earlier this year to have stolen data from around 100 high-profile companies’ Salesforce instances. Salesforce has previously warned that an unnamed known threat actor group was actively scanning for public-facing Salesforce instances and abusing misconfigured guest accounts to break in. Brinks Home is no longer part of the larger Brinks brand, with The Brinks Company telling us it sold the home security arm in 2010. Brinks Home’s parent company, Monitronics, has filed for bankruptcy twice since 2019; for customers’ sake, we hope its physical security services are better than its financial management and infosec. ®

One company's inventive campaign for an unreleased product has become a contest between Anthropic and OpenAI to see which can shout the loudest about its own failures. Readers who tuned in earlier today saw the latest episode in the drama – or sitcom – as Anthropic tried to outdo OpenAI's appropriation of the Mythos marketing playbook and made itself the punchline. Since first teasing Mythos in April, Anthropic has marketed the model through fear – declaring its cybersecurity models too dangerous for public release and offering access only to a select few trusted organizations via Project Glasswing. To its credit, the strategy has paid off. Anthropic has closely associated the Mythos name with cybersecurity, which may explain why OpenAI appeared to borrow its competitor's proven PR strategy last week. OpenAI agents exploited a zero-day to escape their sandbox, leading to the autonomous cyberattack on Hugging Face. The episode duly secured sensational headlines playing on the long-held fear that AI will one day go rogue and take over the world. Anthropic responded this week by lathering on even more clown makeup, squandering an opportunity in the process. The Claude maker sent its models into a testing environment to capture a flag. Their prompts said they had no internet access, but because of what Anthropic called "a misunderstanding" with evaluation partner Irregular, the connection was live. Anthropic's models then followed OpenAI's script: they reached the public internet and attacked systems belonging to outside organizations. This time, three were affected rather than one, the company admitted. In one scenario, Mythos 5 persuaded developers to download a poisoned PyPI package. It was installed on 15 machines, including one at a cybersecurity company that routinely scans such packages for malware. In Anthropic's words: "When that company's scanner installed the package, Claude's hidden code executed. We believe the company's security scanner treated PyPI packages as safe to install, and as a result, Claude was able to exfiltrate the company’s credentials to a collection point it had set up. Claude then used these credentials to access further infrastructure from this company." Worse still, the first of the three incidents occurred in April. Anthropic discovered them only months later, during a retrospective manual review prompted by OpenAI's disclosure. Had it not gone looking, they might never have been discovered, let alone disclosed. There are some caveats. Opus 4.7, the oldest model tested, attacked production systems despite apparently recognizing what it was doing. Mythos 5 recognized that accessing the internet violated its instructions, then reasoned its way into continuing anyway. It was also responsible for publishing the poisoned PyPI package. Only an unnamed research model stopped itself from attacking external organizations. Anthropic also said the models were not running with the production safeguards and monitoring that would normally surround a deployment. Most damningly, Anthropic ran Mythos 5 – the model it had deemed too dangerous for public release – without safeguards in an environment that unexpectedly had internet access. Following OpenAI's admission that it failed so badly in its responsibility to control its technology, Anthropic could have easily spun the story in its favor. You don't have to be fictional tapdancing political PR antihero Malcolm Tucker to see how Anthropic could have used the episode to make its case as the safer, more trustworthy AI company. Instead, realizing its own marketing playbook was being used to help a competitor, it went head-to-head with OpenAI, willingly admitted that it made similar sandbox-based blunders, and disclosed that the results were even more calamitous. Three companies hacked, not just one. So, while the AI biz has attempted to eclipse OpenAI's "rogue agent" story with its own, what's left behind is a new reputation for irresponsible handling of technology. Failed superheroes The incident does not instill a great deal of trust in either Anthropic or OpenAi to safeguard the world from its AI. Dr Ilia Kolochenko, founder of ImmuniWeb and practising cybersecurity and data protection lawyer, likened the two companies to failed superheroes. "While making conclusions would be a bit premature at this point in time, the incidents certainly do not increase confidence in the AI vendor's ability to safely deploy AI, let alone to assure their customers that the so-called frontier models are safe to use," he told The Register. "It is akin to hiring a superhero to protect you but being afraid that the superhero may suddenly go rogue and kill you and your family. Nobody needs such a superhero." Likewise, security pro Jake Williams, VP at HunterStrategy and IANS faculty member, said: "I'm not going to mince words: the major AI labs are negligent in protecting the public from their agents. "We need government regulation now or at the very least a private cause of action with guaranteed punitive damages for agents damaging others." By trying to reclaim a marketing trope that served it well, Anthropic has invited scrutiny of its own safety record and accusations that it is chasing attention above all else. Other experts we spoke to shared the concern that both companies are mishandling their agents, with potentially greater consequences as the systems become more capable. The common thread is recklessness, which Anthropic and OpenAI seem oddly eager to advertise. ®

A week after suspending online banking, CAF Bank still has no timetable for restoring access to its 14,000 UK charity customers. The bank updated customers on Thursday about the outage, which has disrupted payments to staff and suppliers. Little had changed. In a message seen by The Register, CAF Bank said it was not yet able to restore the service safely. As it had earlier in the week, the bank said it detected attempted fraud on some accounts and acted quickly to stop it. Its investigation uncovered a previously unknown vulnerability in the connection between its systems and third-party software. CAF Bank said its technical team was working around the clock with suppliers and external experts on a fix. The Register understands that no timetable has been set for restoring online banking. Earlier this week, CAF Bank CEO Alison Taylor apologized for the disruption. "The core bank is not affected. We are acutely aware of the impact this has on our customers and want this to be fixed as soon as possible, but we cannot restore access to the online service until we are assured the issue is safely resolved," she said. Since The Register first reported the story earlier this week, the BBC has spoken to charities struggling to make essential payments, including payroll. Kevan Hodges, chief executive of Down's syndrome charity 21 Together, told the BBC the outage was "appalling." "People are concerned that wages won't get paid because of this, and that's just stressful when they have bills to pay. My team have wasted days trying to get through to [CAF Bank], but all in vain," he said. Bali Rodgers, chief executive of Safer Communities Alliance, told the BBC the grassroots organizations it represents were slowly losing trust in the bank. CAF Bank also came under fire last year after the introduction of a new banking platform left customers unable to log in or make transactions. The bank later apologized but has not disclosed how much it spent on the system. CAF Bank held £1.45 billion ($1.93 billion) in customer deposits at the end of its 2024/25 financial year. ®

Note to privacy-conscious vandals: If you're going to destroy Flock license plate readers, make sure you also take out the other CCTV cameras in the area that could catch you in your crime. Otherwise, you'll end up like one unlucky Californian. Marcus Bee, 40, was arrested by the Monterey County Sheriff’s Office on Tuesday, accused of joining the ever-growing band of US citizens damaging the controversial cameras popping up across the country. A police report filed this week stated that Bee, of Pismo Beach, was arrested on suspicion of attacking at least three Flock cameras in Lockwood and Bradley. According to the deflock.org website, which maps Flock camera deployments, there are only three of the automated license plate readers (ALPRs) running in the two communities – one in Lockwood and two in Bradley, roughly 25 miles away. Police allege Bee caused thousands of dollars’ worth of damage to “public safety infrastructure.” Monterey County Sheriff’s Office added that Bee was caught after “his actions were captured by other surveillance cameras located nearby,” along with other investigative leads. "This arrest sends a clear message that anyone who intentionally damages public safety equipment will be identified, arrested, and held accountable," said Monterey County Sheriff Tina Nieto. "These cameras have become an invaluable investigative resource that helps us solve crimes, recover stolen vehicles and ag equipment, locate missing persons, and protect our communities. Any attempts to disable these systems will not prevent us from doing our job. “In this case, the suspect's own actions were captured on surveillance cameras, leading directly to his arrest. We will continue to aggressively investigate these crimes and seek prosecution against anyone who targets public safety infrastructure." Bee was jailed with bail set at $30,000. The case follows a similar one in Georgia last week, although local police were unable to identify the suspect(s) behind the two attacks, which involved setting two ALPRs on fire. The two attacks were timed fairly close together, although the phenomenon of inflicting criminal damage onto Flock cameras is something of a long-running trend. Several US police forces have been tasked with arresting alleged Flock vandals, including Jeffrey Scott Sovern, 41, who authorities believe was behind a spate of attacks on ALPR cameras in North Suffolk, Virginia, between April and October 2025. He said, at a hearing in June, that he believed the technology was unconstitutional. Flock cameras: A problem or simply misunderstood? The Monterey County Sheriff’s Office said Flock’s cameras “are an important investigative tool” used to help solve various types of crimes, including missing persons cases, car thefts, and violent crimes such as shootings and homicides where suspect vehicles are involved. Contrary to the opinions held by many, it went on to say, the cameras “are used exclusively to support legitimate criminal investigations,” not to support the US government’s anti-immigration efforts. Likewise, Flock has repeatedly denied offering contracts to agencies such as Immigration and Customs Enforcement (ICE), although reports suggest police were instead carrying out searches on ICE's behalf. Customs and Border Protection (CBP) has also allegedly used Flock data in its own immigration investigations. Other critiques of the technology used by thousands of police departments across the US include problematic abuses, such as police officers using it to stalk romantic interests. The Institute of Justice is aware of at least 26 cases of this behavior, it reported earlier this month, with the majority taking place since 2024. Additionally, the American Civil Liberties Union (ACLU) takes issue with the scale of data gathering by ALPRs. It claims that less than 1 percent of the cars scanned are connected to crime, yet they still have details added to a database, such as vehicle manufacturer, model, color, license plate number, bumper stickers, and scratches. Flock CEO Garrett Langley claimed this week in an interview with The Drive that the company’s cameras were used to solve around 1 million crimes across the US last year. Responding to claims such as Sovern’s – that the cameras are unconstitutional, specifically that they violate Americans’ Fourth Amendment rights – Langley said there are no legal issues, and he doesn’t foresee any arising in the future. Flock’s spokespeople have repeatedly condemned the cases of camera vandalism, highlighting the risk of losing evidence that could be crucial to solving ongoing criminal cases. ®

The Russian espionage crew that turned simply reading an email into a security risk has expanded beyond Zimbra, with Proofpoint saying it's now pulling the same half-click trick against Microsoft Outlook Web Access. Proofpoint says the cyber group it tracks as TA488, or "Laundry Bear," began exploiting CVE-2026-42897, a cross-site scripting flaw in the Outlook Web Access (OWA) component of on-premises Exchange Server, a day before researchers and government agencies exposed the group's abuse of a zero-day in Zimbra Collaboration Suite. Unlike conventional phishing attacks, this one doesn't depend on persuading the victim to follow a link or download a file. If a target opens the booby-trapped message in OWA, the browser executes attacker-controlled JavaScript inside the victim's authenticated mail session. Exchange Online is not affected. According to Proofpoint, TA488 abused the OWA flaw to target government organizations in the US and Europe, along with telecommunications, financial services, hospitality, and aerospace companies. The researchers said the unusually broad campaign may have been intended to hide among the background noise of everyday email traffic rather than the tightly focused operations more commonly associated with espionage groups. "TA488 appears to demonstrate interest in a wide range of sectors while maintaining priorities for intelligence collection against government and defense," Proofpoint said. "Lure themes remain generic and unremarkable, so the target is more inclined to open and skim the email but ultimately overlook it." Instead of dropping conventional malware onto the endpoint, the attackers deploy a browser implant dubbed OWAReaper that lives entirely inside OWA. Proofpoint says it leaves virtually no host artifacts, communicates over two command-and-control channels, supports multiple methods of exfiltrating data, and survives browser restarts, password changes, and even a complete device rebuild because the foothold resides in the compromised mailbox rather than on Windows itself. CVE-2026-42897 isn't making its debut on The Register. Microsoft disclosed the bug in May following reports that attackers were using it in the wild. Proofpoint's latest report fills in more of the picture, showing the activity formed part of a broader espionage campaign rather than isolated exploitation. Proofpoint believes TA488 may actually have been exploiting the flaw as a zero-day, citing attacker infrastructure that dates back to March, roughly two months before Microsoft's out-of-band patch. If accurate, that would suggest the campaign was underway well before defenders knew there was a vulnerability to fix. "If this is the case, the combined improvement of the malware and the exploit development against a harder target in Outlook Web Access signal a leap in capability by TA488," Proofpoint said. Microsoft did not immediately respond to The Register's questions, but if Proofpoint's assessment holds up, TA488 isn't just recycling an old trick. It's refining one that has already proven capable of slipping past one of the oldest pieces of security advice in the book: don't click suspicious links. ®

PWNED Welcome, once again, to PWNED, the weekly column where we show you how not to use your computer or your network. In this week’s fable of fecklessness, we talk about a teacher who had a lot to learn about security. Have a story about someone leaving a gaping hole in their network? Share it with us at pwned@sitpub.com. Anonymity is available upon request. Our story comes courtesy of Kevin Walker, a seasoned IT veteran from the UK. At one point, he was providing his services to a school when he came across the headteacher’s (aka principal’s) laptop. At the bottom of the laptop there was a sticker with the woman’s username and password. Even if they had been complicated, the post-it would have given them away, but in fact, the combination was: Username: headteacher Password: headteacher Using that laptop, a malefactor could have had access to pupils’ personal information, internal conversations, emails, and all kinds of private school files. There could be serious problems for everyone who worked for or attended the school. “A headteacher’s laptop is not just a laptop; it's an entry point to the most sensitive information a school holds,” Walker told us. If a cybercriminal got access, they could effectively break into the school without ever setting foot in the building. This wasn’t the only instance of poor security Walker saw in his time doing IT for schools. He also saw an institution create an Excel file called Passwords.xlsx, then put it on a shared drive that students could get to. As its name suggests, Passwords.xlsx was filled with login credentials that any bad actor could take advantage of. Walker also saw leaver accounts that remained active, a server that had its backup drive permanently plugged in so hackers could potentially wipe the backup as well, a Wi-Fi password written on a whiteboard in reception, and one critical system that users could only access from an ancient laptop. There was also a machine with a “Do Not Turn Off” note posted to it sitting in a corner that everyone was afraid to touch. And, years after Windows XP was no longer the current platform, the school had a CCTV monitor with that ancient OS running on it. And, a supposedly secure server room doubled as a storage closet for stationery and Christmas decorations. Walker told us that, in his experience, the schools he worked with had priorities other than cybersecurity and they didn’t understand its importance. One boss even denied the importance of keeping data safe at all. “We don’t need to worry about cybersecurity. They're only a primary school,” his manager told him when Walker tried to get them to use cloud backups. The problem, Walker opines, is that schools often have to work with outdated gear and the teachers and school administrators have other concerns. His solution: keep it simple. “Make the safe thing the easy thing,” Walker said. “Give staff password managers. Use multi-factor authentication. Review accounts properly. Test backups. Remove shared admin logins. Keep systems updated. Enforce proper passwords and block the ones that have already turned up in data breaches. If a password is already doing the rounds online, it has no business protecting a school system. None of that is as exciting as rolling out a fleet of shiny new iPads, but it works.” ®

UPDATED Watch out for untrusted documents. According to research, an attacker can hide malicious instructions in a Word document that, when included in Copilot for Word’s context, may alter document output and copy the instructions into newly created files that use the affected document as source material, without the victim noticing. Håkon Måløy, a Norwegian data scientist with a PhD in applied AI and ML, publicly disclosed the issue in a blog post Tuesday. Måløy describes the issue in considerable detail while withholding the specific prompt payload, arguing that, because no robust mitigation exists, it would be irresponsible to disclose anything beyond the class of the vulnerability. “To my knowledge, this is among the first public demonstrations of document-borne AI-worm self-propagation through normal workflows in a mainstream commercial productivity suite,” Måløy noted. Måløy said that he has been working with Microsoft since March 2026 on addressing the vulnerability, but after multiple updates to Copilot, this new class of Copilot worm is still viable. Microsoft mitigated the exploit demonstrated by his original proof-of-concept prompt, but Måløy said rewording the payload allowed him to successfully propagate the worm and alter financial data in a target document. Måløy and Microsoft twice delayed public disclosure of the issue, but, after 144 days, he said in his report that people needed to be made aware. “The coordination period agreed with Microsoft has been exhausted, and testing shows that no robust mitigation for the broader vulnerability class is currently available,” Måløy wrote. “Two mitigation attempts, including a model upgrade, did not close the class.” How Copilot propagates a Word worm Måløy explained the worm’s execution with an example involving an employee preparing a financial report for their company. The employee downloads a market analysis from a trusted website to help with the preparation of a financial report in Copilot, unaware that the source had been compromised and the document they downloaded contains hidden malicious instructions. The hidden instructions (inserted as small white text in his proof of concept) tell Copilot to alter figures in the report the employee generates and to copy the worm into the report they create with Copilot. If another employee later adds that report to their own work, the whole process begins again, and documents generated from it also contain the worm, and, as it spreads, it makes tracing the infection to its source extremely difficult. “The attack can therefore continue without further involvement from either the compromised website or the original malicious document,” Måløy said. “The attacker does not need access to the victim’s Microsoft 365 tenant. The attacker only needs to share a malicious document with the victim.” Copilot should use information in documents a user includes in its context for a project without treating instructions embedded in a document as additional prompts, Måløy said, but his research suggests it doesn't always do that. A fundamental flaw Måløy argues that he’s essentially dug up a new type of cross-domain prompt injection attack that abuses a fundamental part of modern LLM architecture. “For AI-assistants to be useful, they often must process emails, documents, webpages, memories, tool outputs, and other information that may be controlled by an attacker,” the researcher said. But if an LLM has to process data in order to determine it contains an attack, the attack could already be influencing that determination. “Relying on the model to detect XPIAs therefore resembles asking an interpreter to execute an untrusted program to determine whether that program is safe to execute,” Måløy asserted. Were Microsoft or some other company to pop another model in front of that model to check for malicious content, it only moves the problem outward, Måløy said, creating a “LLMs all the way down” scenario. “The long-term challenge likely lies in designing systems in which goals and intentions also exist independently of the information being processed,” he said. Until that time, Måløy argues, “any system that integrates an LLM into a trusted workflow today must assume that attacker-controlled content entering the model’s context will result in compromise at some rate.” What can Copilot customers do to reduce the risk? Short of ditching Copilot, there’s not much. “No customer-side remediation fully addresses the issue at the time of publication,” Måløy said, but he does have a few tips. Treat externally sourced documents as untrusted when using them in Copilot, he recommends, and fully review every single document before sending it to Copilot, and fully review any Copilot-generated or edited documents before distributing them. Sheesh - if you’re going to have to actually read that stuff, you might as well just cut Copilot out of the loop and do the thinking yourself. Microsoft has been in touch to confirm the research, but the company's statement doesn't do anything to allay fears this is an unsolved issue. “We have addressed the findings reported by the researcher and thank them for working with us through coordinated vulnerability disclosure. To address this class of risk, we use a defense-in-depth strategy with safeguards that block malicious instructions at multiple points and help keep tasks aligned with users’ requests. We are continuously strengthening these safeguards as the technology and threat landscape evolve. We encourage customers to install the latest updates, use multiple layers of security protection, treat content from unknown sources with caution, and review AI-generated content before using or sharing it.” We also reached out to Måløy, but didn’t hear back before publication. ® Updated at 1841 GMT on July 29 to add Microsoft's statement.

Security researchers at Tenable suspect the Iran-linked faux hacktivist outfit CyberAv3ngers was behind the cyberattack that disrupted more than 30 Minnesota water facilities. Neither state-level nor federal officials have made any claims regarding attribution for the attacks, however, Tenable reckons the operational pattern is consistent with the crew's previous raids, noting the timing relative to recent government warnings. The Cybersecurity and Infrastructure Security Agency (CISA) updated an advisory on Iran-linked attackers targeting programmable logic controllers (PLCs) across critical infrastructure on July 22, four days before Minnesota said the attacks targeted its systems. The advisory warned that Iran-linked hackers were attempting to disrupt operations using tactics previously associated with CyberAv3ngers. Government facilities, water and wastewater systems, and energy providers were among those urged to remain on high alert. What happened in Minnesota? On July 26 and 27, more than 30 community water systems across Minnesota were disrupted by what officials called "a coordinated cyberattack" targeting operational technology (OT). Minnesota IT Services (MNIT), the state's IT agency, said the Department of Health is working with the affected water facilities to ensure public health is maintained. No cities have yet asked citizens to modify the amount of drinking water they consume, per MNIT's latest update. The agency did not offer many other details about the attacks, other than to mention all the different agencies, organizations, and bodies it is working with as part of the investigation. One of the first cities to report issues, Braham, warned that its water reserves were limited in its initial notice. Citizens were asked not to water their lawns or use water for recreational purposes, although the problems were resolved the same day. No such directives were issued in other affected cities. Maple Plain declared a state of emergency, allowing it greater flexibility to coordinate resources, but did not ask residents to adjust their consumption. The same was true in the Twin Cities suburb of Plymouth and in South St. Paul, which both confirmed cyber-related problems on July 27 but did not ask residents to curb water use. "Cyberattacks against critical infrastructure require a coordinated, whole-of-government response," said John Israel, MNIT assistant commissioner and Minnesota CISO. "MNIT is working side by side with our partners to share intelligence, support affected communities, and help utilities restore operations safely while strengthening defenses against future attacks. "This incident demonstrates why Minnesota has invested in strong cybersecurity capabilities and partnerships. Our response worked as intended, enabling agencies at every level of government to rapidly coordinate, contain the incident, and help prevent more serious impacts to critical services." What is CyberAv3ngers? First identified around 2020, CyberAv3ngers is widely believed to be linked to Iran's Islamic Revolutionary Guard Corps (IRGC), specifically its Cyber-Electronic Command division (IRGC-CEC). For the first two years, the group began as a "propaganda persona," as Tenable puts it, claiming disruptive attacks on Israeli infrastructure – claims that were later debunked as fabrications. Its first sustained campaign came in November 2023, when it compromised PLCs at the Municipal Water Authority of Aliquippa, Pennsylvania, defacing them anti-Israel messages. Tenable said CyberAv3ngers compromised at least 75 Unitronics Vision Series PLCs across the United States, Israel, the United Kingdom, and Ireland as part of the campaign. They did so by exploiting default passwords. Between 2024 and 2025, the crew developed the IOCONTROL malware kit, built for attacks on OT and Internet of Things (IoT) devices. OpenAI said in 2024 that the group's members used ChatGPT in the development process. CyberAv3ngers stepped up its activity in 2026, targeting US critical infrastructure through Rockwell Automation/Allen-Bradley PLCs from March onward. CISA's July 22 update added Schneider Electric and Siemens equipment to the list of potential targets. In some cases, the attacks - which targeted multiple critical infrastructure sectors - disrupted operations at affected facilities, federal officials said, though they offered no specifics on what those disruptions entailed. CyberAv3ngers is known for targeting small water and municipal facilities, which experts believe are among the lowest-hanging fruit in US critical infrastructure. Many small and rural facilities lack dedicated cybersecurity resources. Tenable said some operators manage OT environments using remote-access software such as TeamViewer and AnyDesk or leave their PLCs exposed to the web. "These access methods bypass enterprise security controls entirely, creating an attack surface that is invisible to conventional security monitoring," Tenable said. Poor segmentation between IT and OT environments can also allow a single intrusion to spread across much of the network. ®

The US government has decided to effectively ban the sale of advanced robots made in other nations. The decision trickled out over two days with publication of a National Security Determination [PDF] and an update [PDF] to the list of banned devices set by the Federal Communications Commission (FCC). The national security document observes “Advanced robotic devices will be critical to creating efficiencies in our economy, dominating on the battlefield, and securing our homeland” and notes that modern bots are now constantly connected to networks “which creates broad attack surfaces and leaves them vulnerable to data exfiltration, remote disruption of the physical robot, and dependencies on unsecure over the air updates.” One example of those vulnerabilities mentioned in the document is the UniPwn flaws that made it possible for attackers to take over humanoid robots made by Chinese company Unitree. “If the United States continues to rely on foreign sources of advanced robotic devices and critical components, it will subject the parts of the U.S. economy and national security enterprise that are reliant on these robots to the whims of foreign entities that could disrupt or degrade the supply chains at a time of their choosing,” the document states. To respond to those threats, the FCC decided the foreign-made advanced robotic devices belong on its Covered List of products for which imports are banned because they pose an unacceptable risk to the national security of the United States and its residents. The regulator offered a single exception: if the Department of War vouches for a device, it can have it. Foreign-owned companies that make their bots in America are also exempt, an important exemption because one of the leading robot-makers is Boston Dynamics – a company backed by the USA’s DARPA that is now majority-owned by South Korea’s Hyundai, but continues to manufacture its machines stateside. The decision does, however, apply to all future foreign-made devices. Vendors of clankers already approved for sale in the USA can continue to import them, and users are also free to use any bots they already own. But the intent of the documents is clear: from now on, only robots made in America are welcome in America. One entity that stands to benefit from this decision is Tesla, which Elon Musk claims will one day produce one million humanoid robots a year. In true Muskian style he has also said Tesla will go into “high production” of the bots in 2026, but there’s no evidence of that happening although the occasional trillionaire did recently show off the production line for Tesla’s “Optimus” bot. ®

UPDATED We now know how OpenAI's models broke out of their cages to attack Hugging Face. The rogue models found zero-day vulnerabilities in JFrog’s universal binary repository manager Artifactory around the time they escaped, according to JFrog CTO Yoav Landman. While Landman wouldn't confirm that these flaws were the zero-days that OpenAI’s models found and exploited, ultimately allowing them to breach the massive model mart, OpenAI later admitted the connection. "To gain Internet access, the models identified and exploited a previously unknown zero-day vulnerability in Artifactory⁠ package registry cache proxy," OpenAI added to a blog post on the topic Tuesday. "We disclosed this vulnerability, along with other Artifactory vulnerabilities our models identified as part of our review, to the vendor." Landman says OpenAI's models discovered the Artifactory zero-days during a security evaluation. The AI giant notes the incident occurred while its models were being evaluated on the ExploitGym benchmark. “During a security evaluation, OpenAI’s models identified previously unknown zero-day vulnerabilities in self-hosted Artifactory installations that could be exploited to gain unintended internet access,” Landman said on Monday. JFrog Artifactory is a central platform that organizations use to store and distribute all the software artifacts across their supply chains. It supports more than 60 package formats including Docker, Maven, npm, PyPI, Helm, and AI/ML models. OpenAI “responsibly and immediately” disclosed the vulnerabilities to JFrog, Landman continued. “Our security team treated the report with the urgency it deserved, as a genuine zero-day unknown to the world, and moved accordingly. We developed, validated, and released a fix for all JFrog customers, self-hosted and cloud alike.” On Monday, JFrog released the fixed versions, and credited OpenAI researchers for reporting at least eight of the now-patched Artifactory vulnerabilities: CVE-2026-65617, CVE-2026-65925, CVE-2026-65921, CVE-2026-65923, CVE-2026-66018, CVE-2026-66014, CVE-2026-66015, and CVE-2026-65924. JFrog's admission comes about a week after OpenAI said two of its models, GPT-5.6 Sol and a second pre-release model, escaped their testing sandbox during a security evaluation designed to test their cyber capabilities. During this test, the models found a way to access the open internet, then broke into Hugging Face and accessed private information and stole some credentials. “While operating in our sandboxed testing environment, our models spent a substantial amount of inference compute finding a way to obtain open Internet access, in pursuit of solving the evaluation problem,” OpenAI said on July 21. In a July 28 update, the house of Altman admitted that the JFrog 0-days were the cause. In the same update, OpenAI admitted that its models had breached other services. "We have been finding a small number of cases where the models identified and used publicly exposed credentials at the account-level on other publicly-available services. This includes four accounts on four services as part of the Hugging Face incident (and a few accounts accessed as part of other evaluations)," OpenAI wrote. ® Updated on July 29 with additional details from OpenAI, admitting that the JFrog 0-days led to the breach.

Two agentic bug-hunting systems from Microsoft and Google-owned Wiz show that when it comes to finding and remediating software vulnerabilities, at least two models’ minds work better than one - and Wiz tells us it’s adding a third. Wiz on Monday said Project Atlas, its bug-hunting AI agent, bested Anthropic’s Mythos Preview and OpenAI’s GPT-5.5 Cyber with its vulnerability-analysis skills, achieving a 90.9 percent success rate on CyberGym, and uncovering more than 200 zero-day security holes in widely used open-source code. Meanwhile, Microsoft boasted its MDASH bug-hunting harness scored a 95.95 percent success rate on CyberGym, also beating Mythos, Gemini and GPT on the same benchmark for evaluating how well AI systems find real vulnerabilities in the code. For comparison, OpenAI’s GPT-5.5 Cyber scored 85.6 percent on CyberGym, and its GPT-5.6 Sol scored 83.6 percent. Anthropic’s Mythos 5 reproduced the target vulnerability on 83.8 percent of CyberGym challenges. And Google’s Gemini 3.5 Flash Cyber in CodeMender achieved an 83.2 percent success rate. The secret to both Atlas and MDASH’s success, according to the vendors, is that they use the right model for the right security job. Atlas uses Claude Opus 4.6 with GPT-5.5, Nir Ohfeld, head of vulnerability research at Wiz, told The Register. “We're now working to incorporate Gemini, which is well timed given Wiz's recent work with DeepMind on Gemini Flash Cyber,” he added. Microsoft’s MDASH - a combination of red-team agents that find and simulate real, exploitable vulnerabilities and attack paths, and green-team agents that remediate the issues - combines MAI-Cyber-1-Flash, based on Microsoft AI (MAI)’s internally developed MAI-Thinking-1 reasoning model, and GPT-5.4. MAI-Cyber-1-Flash is designed to handle up to 90 percent of all tasks, with MDASH detecting, patching, and validating vulnerabilities before handing the remaining 10 percent of more complex tasks to the larger GPT-5.4. “We were able to take an off-the-shelf model, within our harness, a multi-agent and multi-model implementation, and we achieved the best results you could have,” Hayete Gallot, executive vice president of Microsoft Security, said on Monday. Atlas isn’t commercially available yet - it’s used internally, and stems from Wiz’s efforts to understand how frontier models can be used for advanced code scanning. But it’s proof that “no single model is best at everything, and none stays state of the art for long,” Ohfeld and fellow Wiz kid Yuval Avrahami wrote in a Monday blog. The cloud security biz evaluates every new model using its internal benchmarking tool, Cyber Model Arena, which scores each one on its success at completing various security-investigation tasks: threat modeling, hunting, validation, and proof generation. “The results are rarely uniform: the model that reasons best through a complex exploit chain is often not the one that triages most precisely,” the duo wrote. “Atlas routes each stage to whichever model wins on that task.” In addition to doing a better job of finding and fixing vulnerabilities, a multi-model system also saves customers’ money, according to Microsoft and Wiz. Combining its much smaller, in-house model with GPT-5.4 halves customers’ costs, according to Mustafa Suleyman, CEO of Microsoft AI. “As the models hand off between each other, they are not just able to deliver better performance than all of the other models combined, they do so at 50 percent of the cost,” he said on Monday. And while “each new generation of models expands what is possible,” they are also expensive, Ohfeld told us. “We have also learned that pointing a frontier model at a codebase once is not a sustainable security strategy: deep scans are expensive, their results become stale as code changes by the minute, and a point-in-time analysis cannot provide the continuous coverage organizations need across every repository,” he said. In fact, the real question for code security shouldn’t be which model a scanner uses, Ohfeld added. It’s this: “How does your system take advantage of the best model available today, continuously and economically, and what continues to work when a better one arrives,” he said. “That is the bet behind Atlas: frontier-model depth where expert reasoning is required, an architecture that improves as models evolve, and rigorous validation so every finding arrives with evidence, not just a plausible answer.” ®

Ahead of DEF CON 2026 opening its doors in Las Vegas next week, conference organizers said they have imposed a ban on “Meta-style glasses with recording capabilities.” Statements made via its social media channels went on to say that no exceptions will be made for those using the devices with prescription lenses. “Be sure to pack non-violating eyewear if you need them,” DEF CON said, before directing delegates to the conference’s official photo policy. That policy has not been updated since 2023, predating the recent growth of camera-equipped eyewear developed by Meta with EssilorLuxottica under its Ray-Ban and Oakley brands. It states that public photography is permitted but with several caveats that essentially prohibit capturing the image of anyone, except on-stage speakers, unless the photographer obtains consent from the subject(s). EFF director of cybersecurity Eva Galperin welcomed the decision, saying: “Love to see a ‘no pervert glasses’ policy at DEF CON.” The conference’s reminder follows similar moves from other major organizers and promoters. Monopoly Events banned the same style of smart glasses from its shows, which include UK Comic Cons, in recent days. “After a consultation period, we can no longer permit the wearing of any recording device at our events,” it stated last week. “The vast majority of the talent and agents at our events felt that these were a violation of their privacy and were damaging and spoiling the interactions at the tables. “Several expressed concerns that they might not continue to attend in-person events if they were being recorded in secret. “We had similar feedback from event attendees, many of whom are not comfortable with the knowledge that other convention goers might be filming them without consent.” The event management company said that anyone caught wearing glasses with recording capabilities may be asked to leave the show and forfeit any unfulfilled autographs or photographs they had purchased. Similarly, Scottish ferry operator CalMac temporarily suspended unplanned visits to ships' bridges of its ships after a passenger wearing recording glasses made crew and passengers feel uncomfortable while filming during a crossing in June, the Ayrshire Weekly Press reported. Growing privacy concerns Originally pitched as a convenient tool for recording everyday moments without needing to occupy one’s hands with a phone or camera, Meta’s smart glasses quickly attracted a mixed reception. Old fans of the Google Glass project were enamored by Meta’s take on the concept, more than ten years after the Chocolate Factory debuted its chunkier, costlier wearable, which was swiftly axed after two years. Meta’s glasses, and crucially their recording capabilities, are considerably more clandestine than Google Glass', appearing to many as a normal set of specs. Only when you inspect the frame from a much closer perspective does the embedded camera become more apparent. The devices have become associated with creepy behavior. A quick search for “Meta glasses privacy violations” will throw up countless examples of questionable conduct from glasses-wearers, typically reported by women and children. While photography and videography in public spaces are widely permitted for casual use, smart recording glasses make the activity much more discreet. The Register has heard that such devices can be paired with unsophisticated apps to dox passersby in seconds, and have inspired separate projects to alert Android users to nearby glasses-wearers using Bluetooth signals. Meta is also facing scrutiny from the UK’s data protection watchdog, including over cross-border data flows. The watchdog’s questions follow reports originating in Sweden that Kenya-based contractors reviewing footage from wearers were exposed to some of their more private moments. Human reviewers based in Kenya, who are tasked with labelling images and video to help train Meta’s AI, have reportedly reviewed captures taken from toilet visits, wearers changing their clothes, and users engaging in conversations that revealed alleged wrongdoing. Meta has routinely defended its wearables amid privacy concerns, saying that when they are recording, a light on the frame illuminates to indicate that the camera is active. Further, attempts to cover or otherwise tamper with this light result in the glasses refusing to capture images. The Register contacted Meta for its take on the recent bans of its devices at conferences and will update this article if we hear back. ®

Anthropic's Project Glasswing may have uncovered tens of thousands of potential security flaws, but new research suggests AI-assisted vulnerability discovery has yet to produce the wave of real-world attacks many expected. In research shared with The Register, VulnCheck analyzed 1,061 publicly attributed AI-assisted vulnerability discoveries from Anthropic's Project Glasswing and the Berkeley Vulnerability Research Initiative, then cross-referenced them against its Known Exploited Vulnerability (KEV) database. The result: just 14 vulnerabilities, or 1.3 percent, have been confirmed as exploited in the wild, almost identical to the rate across all vulnerabilities in VulnCheck's dataset. That's a far cry from the narrative that frontier AI is dramatically tilting the balance in attackers' favor by churning out instantly weaponizable bugs. Instead, the data suggests that AI is currently better at increasing the volume of vulnerabilities researchers can uncover than at increasing the proportion that attackers actually exploit. The report takes particular aim at Anthropic's much-publicized Project Glasswing, unveiled in April with warnings that AI-assisted vulnerability discovery could allow attackers to hijack systems, disrupt operations, or steal data. Claude Mythos may have identified 23,019 vulnerability candidates, but there's remarkably little public evidence showing what became of most of them. VulnCheck notes that only 126 have been published as CVEs, that just one has been confirmed exploited in the wild, and that Anthropic's public disclosure record has seen little movement since Project Glasswing launched. But that doesn't mean AI-assisted vulnerability research has failed, according to Patrick Garrity, security researcher at VulnCheck. "AI-assisted vulnerability discovery clearly has value for both attackers and defenders," Garrity wrote. "The data does not suggest that AI-discovered vulnerabilities are inherently more likely to be exploited than those found through traditional methods." Instead, he argues, AI is simply helping researchers discover more flaws, giving defenders an opportunity to patch them before criminals get there. Garrity stopped well short of declaring the threat overblown forever, but he did suggest some of the rhetoric has outpaced reality. "The data so far, including Anthropic's own stalled disclosure ledger, suggests that AI-assisted vulnerability discovery and frontier capabilities have been overhyped relative to the evidence available today," he wrote. "That doesn't mean the risk is imaginary. It means the impact has been real but modest." Meanwhile, attackers haven't exactly been sitting idle. VulnCheck identified 495 known exploited vulnerabilities during the first half of 2026, with content management systems accounting for roughly one-third of them and network edge devices remaining a firm favorite. AI products themselves are also becoming an increasingly attractive target, as attackers look beyond using AI and start hunting for weaknesses in the rapidly expanding AI software stack. In other words, AI may be changing vulnerability research, but it hasn't yet produced the exploitation apocalypse some predicted. ®

CAF Bank, which serves 14,000 charities, has suspended online banking as it fixes a vulnerability in how third-party software connects to its portal. The outage has left some organizations struggling to run payroll after being cut from their accounts. In a message to customers, seen by The Register, the Charities Aid Foundation-owned bank confirm online services had been unavailable since July 24 and will remain so until further notice. The bank's communications say the decision resulted from reports of suspicious activity on some customer accounts. It said the bank detected the problem early on and notified customers of any attempted fraud. Following an investigation, CAF Bank identified a previously undetected vulnerability in the connection between third-party software and the online banking portal. It is working with its technology partner on a fix. The bank assured customers that its core banking services were not affected and that money held in their accounts was safe. However, it was making changes to the online service. In a statement, CEO Alison Taylor said: "We have informed CAF Bank customers that the online banking service will be unavailable until further notice. I am very sorry for the disruption and understand the frustration this can cause for our customers. "We are working with external experts to fix an issue we identified with third-party software related to our online banking portal. The core bank is not affected. We are acutely aware of the impact this has on our customers and want this to be fixed as soon as possible, but we cannot restore access to the online service until we are assured the issue is safely resolved. "We are still able to support on the phone, and we are prioritizing time-sensitive payments such as payroll." She declined to comment on whether the bank will compensate customers. Last year, CAF Bank came under fire from customers who were unable to log in or make transactions when a new banking platform was introduced. The bank later apologized to customers experiencing difficulties with the service. At the time, a CAF Bank spokesperson said it was focused on supporting its customers through the transition and helping them with the new online banking service. "The vast majority of our customers are online, with thousands of payments being made and received every day. We are sorry for the disruption and waits on calls that some of our customers have experienced following the launch." The bank has not disclosed how much it has spent on the platform. It held £1.45 billion ($1.93 billion) in customer deposits at the end of its 2024/25 financial year. ®

A critical flaw in Arista's VeloCloud Orchestrator has gone from zero to KEV in short order, with the networking giant confirming attackers are already exploiting it. The vulnerability, tracked as CVE-2026-16812, carries a maximum CVSS score of 10.0 and affects VeloCloud Orchestrator On-Prem, the self-hosted version of the software that enterprises use to centrally manage VeloCloud software-defined wide area networks (SD-WANs) connecting branch offices, datacenters, and clouds environments. According to Arista's security advisory, the flaw is an OS command injection vulnerability that allows an unauthenticated remote attacker to reach privileged internal functionality that was never meant to be exposed externally. Worse, Arista says the on-premises orchestrator is exposed by default, with no configuration capable of removing that exposure entirely. Exploitation requires access to the web interface but no credentials. Until administrators can patch, Arista recommends restricting that interface to trusted management networks and blocking IP addresses associated with observed attacks. "Successful exploitation may compromise the confidentiality, integrity, and availability of the orchestrator and data managed by the orchestrator," Arista warned. "Compromises to the VCO platform may allow attackers access to the VeloCloud Edge devices as well." Arista published three IP addresses observed conducting attacks, but otherwise kept its cards close to its chest. The company hasn't said who's exploiting the bug, when the attacks began, or how many customers have been affected, and didn't immediately respond to The Register's questions. Even without those details, the admission of in-the-wild exploitation was enough for CISA to add the flaw to its Known Exploited Vulnerabilities (KEV) catalog. The list is reserved for bugs with evidence of real-world abuse, and while the associated directive applies only to US federal civilian agencies, plenty of private sector security teams use KEV to decide which patches can't wait. The issue affects only on-premises deployments. Customers using Arista's hosted or dedicated VeloCloud Orchestrator service had already been patched before the advisory was published, the company said. Fixes are available in VeloCloud Orchestrator versions 5.2.3.14, 6.1.3.4, 6.4.2.4, and 7.0.0.1. Arista urged customers running earlier releases to upgrade immediately. Arista is far from the first vendor to issue a patch after attackers had already begun exploiting the flaw. Over the past year, a steady stream of networking gear, VPNs, firewalls, and other edge-facing enterprise software has followed the same pattern: by the time customers learn there's a problem, somebody else has already proved it's worth exploiting. ®