The Register

Even limited voter rolls can be linked to identify people, research shows

Your voter data could be used against you. A foreign intelligence service that wished to identify the family members of deployed military personnel could do so by cross-referencing public voter record data and social media posts.…

Information security agencies from the nations of the Five Eyes security alliance have co-authored guidance on the use of agentic AI that warns the technology will likely misbehave and amplifies organizations’ existing frailties, and therefore recommend slow and careful adoption of the tech. The agencies delivered that position last Friday in a guide titled Careful adoption of agentic AI services [PDF] that opens with the observation that “Agentic artificial intelligence (AI) systems increasingly operate across critical infrastructure and defense sectors and support mission-critical capabilities,” making it “crucial for defenders to implement security controls to protect national security and critical infrastructure from agentic AI-specific risks.” The thrust of the document is that implementing agentic AI will require use of many components, tools, and external data sources, creating an “interconnected attack surface that malicious actors can exploit.” “Consequently, every individual component in an agentic AI system widens the attack surface, exposing the system to additional avenues of exploitation,” the document warns. To illustrate the risks agentic AI poses, the document offers the example of an AI agent empowered to install software patches that is thoughtlessly given broad write access permissions, with the following unpleasant results: Here’s another nasty agentic mess the document uses as a warning: An organization deploys agentic AI to autonomously manage procurement approvals and vendor communications, and gives the agent access to financial systems, email and contract repositories; This user only considers permissions for the agent when deploying it; Over time, other agents rely on the procurement agent’s outputs and implicitly trust its actions; A malicious actor compromises a low-risk tool integrated into the agent’s workflow and inherits the agent’s over-generous privileges; The attacker uses that privileged access to modify contracts and approve unauthorized payments, and evades detection by creating faked audit logs that don’t trip alerts. Australia’s Signals Directorate and Cyber Security Centre (ASD’s ACSC) contributed to the document, working with the USA’s Cybersecurity and Infrastructure Security Agency (CISA) and National Security Agency (NSA), the Canadian Centre for Cyber Security (Cyber Centre), the New Zealand National Cyber Security Centre (NCSC-NZ) and the United Kingdom National Cyber Security Centre (NCSC-UK). The document contains more scary stories, then lists 23 different risks and over 100 individual best practices to address them. Much of the advice targets developers who deploy AI, but the authors also urge vendors to ensure they test their wares thoroughly and ensure their products “fail-safe by default requiring agents to stop and escalate issues to human reviewers in uncertain scenarios.” The document also urges security practitioners and researchers to spend more time contemplating AI. “Threat intelligence for agentic AI systems is still evolving, which can introduce significant security gaps,” the document warns, because resources like the Open Web Application Security Project and MITRE ATLAS currently focus on LLMs. “As a result, some attack vectors unique to agentic AI may not be fully captured or addressed.” Given the huge to-do list for anyone creating agentic AI, or contemplating its use, the document argues for very cautious adoption. “Organisations should therefore approach adoption with security in mind, recognizing that increased autonomy amplifies the impact of design flaws, misconfigurations and incomplete oversight,” the document concludes. “Deploy agentic AI incrementally, beginning with clearly defined low-risk tasks and continuously assess it against evolving threat models.” “Strong governance, explicit accountability, rigorous monitoring and human oversight are not optional safeguards but essential prerequisites. Until security practices, evaluation methods and standards mature, organisations should assume that agentic AI systems may behave unexpectedly and plan deployments accordingly, prioritizing resilience, reversibility and risk containment over efficiency gains.” ®

Prioritize resilience over productivity, say CISA, NCSC and their friends from Oz, NZ, Canada

Information security agencies from the nations of the Five Eyes security alliance have co-authored guidance on the use of agentic AI that warns the technology will likely misbehave and amplifies organizations’ existing frailties, and therefore recommend slow and careful adoption of the tech.…

Britain's cyber agency is warning that AI-fuelled bug hunting is about to flush out years of buried flaws, leaving defenders scrambling to keep up. In a blog post on Friday, Ollie Whitehouse, CTO of the UK's National Cyber Security Center, said organizations should brace for a looming "patch wave," driven by a backlog of weaknesses now being exposed faster than many teams can realistically fix them. "All organizations have 'technical debt'; a backlog of technical issues – that is both expensive and time-consuming – as a result of prioritising short-term gains over building resilient products," Whitehouse wrote.  "Artificial Intelligence, when used by sufficiently-skilled and knowledgeable individuals, is showing the ability to exploit this technical debt at scale and at pace across the technology ecosystem," he added. The result, according to NCSC, is likely to be a "forced correction" as those weaknesses are uncovered and addressed in bulk. That warning lands just as vendors roll out tools built to do exactly that. Models like Anthropic's Claude Mythos and OpenAI's GPT-5.5-Cyber promise to find and fix bugs before attackers do, but the same capability also lowers the barrier to finding them in the first place. "We are expecting an influx of updates to address vulnerabilities across all severities, and expect a number to be critical," Whitehouse wrote. The cyber agency is urging teams to get ahead of the incoming flood by shrinking their exposed footprint. "All organizations must take steps to identify and minimise their internet-facing (and other externally-exposed) attack surfaces as soon as is possible," Whitehouse said, adding that defenders should "prioritise technologies on your perimeter and then work inwards." Even then, patching alone will not be enough; Whitehouse notes that unsupported or end-of-life systems may need to be replaced altogether. "Prepare to patch quickly, more often, and at scale," is the message from the NCSC. In practice, that means a lot more fixes landing at once, and a lot less time to get them done. ®

Britain's cyber agency says the bill for years of technical shortcuts is coming due, and it's arriving all at once

Britain's cyber agency is warning that AI-fuelled bug hunting is about to flush out years of buried flaws, leaving defenders scrambling to keep up.…

CISA has added a critical cPanel bug to its known-exploited list, confirming that attackers are already poking holes in one of the internet's most widely used hosting stacks. The vulnerability, tracked as CVE-2026-41940, carries a near-worst-case CVSS score of 9.8 and affects all supported versions of cPanel and Web[Host Manager (WHM) released after version 11.40, along with WP Squared, a WordPress management layer built on top of the same platform. In plain terms, a successful exploit can hand over full control of the server. The US government's cybersecurity agency added the flaw to its Known Exploited Vulnerabilities catalog on Thursday, confirming attackers are not waiting around. By the time cPanel shipped a patch on Tuesday, exploitation was already underway. Hosting provider KnownHost has been more explicit about what that looked like in practice, warning customers it had seen successful exploitation attempts before any fix was available. In a Reddit post, the company's CEO, Daniel Pearson, said the provider had "seen execution attempts as early as 2/23/2026" and urged users to restrict access and assume systems could already be compromised if left unpatched. Another hosting provider, Namecheap, says it temporarily blocked access to cPanel and WHM, effectively slamming the door shut until fixes were ready. It has since begun rolling out updates. There are also early signs of what those attackers are up to once they get in. A small business owner posting on Reddit said their company had been hit by ransomware after running what they described as a fairly standard cPanel setup, adding that their hosting provider appeared to be struggling under the weight of the incident. The attackers, they said, demanded $7,000 to unlock systems. The claim is anecdotal, but if it holds up, it suggests this bug is already being used by criminals to lock up systems, not just lurk quietly or skim data in the background. It's not yet known how many organizations have been impacted by the vulnerability, but security firm Rapid7 used Shodan to identify roughly 1.5 million internet-exposed cPanel instances.  cPanel underpins hosting for tens of millions of sites, many run by small outfits that rely on providers to handle security. For them, "patch now" often means "wait and hope," which is not a great place to be when a near-max severity bug is already being weaponized. ®

Exploitation was underway before patches landed, at least one victim reports ransomware demand

CISA has added a critical cPanel bug to its known-exploited list, confirming that attackers are already poking holes in one of the internet's most widely used hosting stacks.…

OpenAI is lining up a limited release of its new GPT-5.5-Cyber model to a handpicked circle of "cyber defenders," just weeks after taking a swipe at Anthropic for doing almost exactly the same thing. CEO Sam Altman said in a post on X that the rollout will begin "in the next few days," with access restricted to a group he described as trusted defenders working to secure critical systems.  "We will work with the entire ecosystem and the government to figure out trusted access for cyber," he wrote, adding that the goal is to "rapidly help secure companies and infrastructure." GPT-5.5-Cyber is built to spot flaws before anyone else abuses them. OpenAI says it can pentest, find bugs, exploit them, and tear apart malware, but as we have already seen, tools that break systems rarely stay in the right hands for long. OpenAI's announcement comes just weeks after Anthropic rolled out its own cyber-focused model, Claude Mythos, to roughly 50 organizations under tight controls, saying it would never be made publicly available – and Altman was not impressed.  As reported by TechCrunch, he took aim at what he framed as exclusivity dressed up as caution during an appearance on the Core Memory podcast.  "There are people in the world who, for a long time, have wanted to keep AI in the hands of a smaller group of people," he said. "You can justify that in a lot of different ways." He went further, likening the approach to selling fear. "We have built a bomb, we are about to drop it on your head. We will sell you a bomb shelter for $100 million." Now OpenAI is, if not building the same shelter, at least checking IDs at the door. Independent testing suggests the model is not just marketing fluff. The UK's AI Security Institute said this week that GPT-5.5-Cyber is "one of the strongest models we have tested on our cyber tasks," and noted it is only the second system to complete one of its multi-step attack simulations end to end.  It may be pitched as protection, but when the tools can both break and fix systems, the difference often comes down to who gets there first. ®

Altman's crew now doing the same gatekeeping it recently mocked

OpenAI is lining up a limited release of its new GPT-5.5-Cyber model to a handpicked circle of "cyber defenders," just weeks after taking a swipe at Anthropic for doing almost exactly the same thing.…

Canonical says its web infrastructure is under attack after a pro-Iran hacktivist group instructed its members to target the open source giant. "I can confirm that Canonical's web infrastructure is under a sustained, cross-border Distributed Denial of Service (DDoS) attack" a Canonical spokesperson told The Register.  "Our teams are working to restore full availability to all affected services. We will provide updates in our official channels as soon as we are able to." Known best for managing the development of Ubuntu, the distro's main website is down at the time of writing, and has been for several hours. The hacktivist group The Islamic Cyber Resistance in Iraq, aka 313 Team claimed responsibility for the 503 errors Ubuntu's website was returning on Thursday evening, announcing via its Telegram channel that the attack was scheduled to persist for four hours. More than 12 hours later, the attack continues to disrupt Ubuntu's main website and many of its subdomains, although some, including its Archive and Discourse pages, remain up and running. 313 Team sent a follow-up message to its Telegram group, directed at Canonical, which indicates the group is veering away from hacktivism and toward full-on extortion: "There is a simple way out. We have emailed you with our Session Contact ID. If you fail to reach out, we will continue our assault. You are in an awful position, don't be foolish." The service disruption at Ubuntu means users cannot download any versions of its distros through the usual channels, nor can they log into their Canonical accounts. Canonical promised to provide regular updates when it has new information to share. 313 Team has claimed responsibility for similar DDoS attacks on the likes of eBay's Japan and US divisions, as well as BlueSky in just the past month alone. Why the group is targeting London-based Canonical remains unclear and no reason was given via its Telegram channel. It is presumably because Ubuntu is one of the most popular Linux distros. ®

313 Team tells Canonical: pay up or the packets keep coming

Canonical says its web infrastructure is under attack after a pro-Iran hacktivist group instructed its members to target the open source giant.…

The Home Office has increased the annual value and overall duration of its new passport production contract, increasing it to a total of £576 million as it starts a third round of engagement with suppliers. The department’s first engagement notice for the Provision of Passport Manufacturing and Personalisation Services contract last July included an estimated total value of £360 million including VAT over 10 years or £36 million a year. The version published on 24 April increases the total value to £576 million including VAT over 12 years or £48 million a year. The Home Office has also pushed back the contract’s start date from September 2027 to August 2028, as well as postponing the publication of the full tender notice from June to November this year. The latest version says that HM Passport Office issues about eight million passports annually, up from seven million in the first notice, although this would not fully account for the increased annual value. The Home Office’s current passport production contract with Thales (which bought the winning bidder Gemalto) started in April 2018, with an estimated value of £262 million over 11.5 years or £22.8 million a year. It ends on 30 September 2029. As well as physical production of passports and other travel documents, the new supplier will have to personalize them with data including biometrics. It may also need to produce digital travel credentials and make provision “for crypto technologies and contingency solutions.” Potential suppliers will have the chance to ask questions after completing a non-disclosure agreement at an online event on 18 May. The Home Office disclosed that it will pay IBM £5.88 million including VAT for software licenses and support services to operate and maintain its biometric systems between 1 May 2026 and 30 April 2028. The department is awarding the contract directly without competition "as the required software and support services are proprietary to IBM and embedded within existing live systems, with no reasonable alternative supplier without disproportionate technical difficulties." ®

Start date pushed back a year, annual cost up a third, and UK's now handing out eight million passports a year

The Home Office has increased the annual value and overall duration of its new passport production contract, increasing it to a total of £576 million as it starts a third round of engagement with suppliers.…

The wave of supply chain attacks aimed at security and developer tools has washed up more victims, namely SAP and Intercom npm packages, plus the lightning PyPI package. The newly compromised packages as of Thursday include intercom-client@7.0.5 (according to Google-owned Wiz) and intercom-client@7.0.4 (says supply-chain security firm Socket) and lightning@2.6.2 and 2.6.3. Attackers infected all versions with the same credential-stealing malware that, on Wednesday, poisoned multiple npm packages associated with SAP's JavaScript and cloud application development ecosystem. The SAP-related compromise is a Shai-Hulud-worm style campaign that calls itself Mini Shai-Hulud. So far, these SAP-related npm packages include: mbt@1.2.48 @cap-js/db-service@2.10.1 @cap-js/postgres@2.2.2 @cap-js/sqlite@2.2.2 Collectively, these four packages receive about 572,000 weekly downloads and are widely used by developers building cloud applications. SAP did not answer The Register's questions about the compromise and instead sent us this statement: "A security note is published and available for SAP customers and partners." The note is only accessible to logged-in customers. These latest offensives are called "Mini Shai-Hulud worm” attacks because of similarities to the earlier self-propagating Shai-Hulud malware that targeted npm packages. Both Wiz and Socket attributed the SAP compromise to TeamPCP – the cybercrime crew linked to the earlier Checkmarx, Bitwarden, Telnyx, LiteLLM, and Aqua Security Trivy infections. The two security shops also note that the Thursday attacks on the Intercom and lightning packages appear to contain the same malicious code seen in the SAP operation. Here's what has happened in the world of supply-chain attacks over the past 48 hours. SAP-related npm packages On April 29, TeamPCP compromised four official npm packages from the SAP JavaScript and cloud application development ecosystem and published the poisoned releases between 09:55 and 12:14 UTC. The compromised packages contain malicious preinstall scripts set to execute automatically on every npm install, and run attacker-controlled code before any application code runs. This new campaign deploys a multi-stage payload that steals developer secrets, self-propagates, encrypts all the stolen goods, and then exfiltrates the now-locked secrets into a new GitHub repository under the victim's own account. "The second-stage payload is a credential stealer and propagation framework designed to target both developer environments and CI/CD pipelines," the Wiz kids said on Thursday. "It collects sensitive data including GitHub tokens, npm credentials, cloud secrets (AWS, Azure, GCP), Kubernetes tokens, and GitHub Actions secrets – leveraging advanced techniques such as extracting secrets from runner memory. Exfiltration occurs via public GitHub repositories, where it posts encrypted payloads. Additionally, the malware includes propagation logic to infect additional repositories and package distributions." Plus PyPI package lightning Then on Thursday, an additional package was poisoned to execute credential-stealing malware on import. Up first: PyPI package Lightning versions 2.6.2 and 2.6.3. Lightning is a widely used deep learning framework for training and deploying AI products. Developers download it hundreds of thousands of times every day. "The obfuscated JavaScript payload contains many similarities to the Shai-Hulud attacks, overlapping in targeted tokens, credentials and obfuscation methods. Socket also identified signs that router_runtime.js both poisons GitHub repositories and infects developer npm packages," according to Socket, which also published a separate Mini Shai-Hulud supply-chain campaign page that it updates as new information comes to light. And Intercom's npm package Also on Thursday: Socket and Wiz sounded the alarm on a new compromise of the intercom-client npm package. Intercom is a customer communications platform, and intercom-client is a widely used official SDK for Intercom's API. It sees about 360,000 weekly downloads, and npm lists more than 100 dependent projects. However, as Socket notes, the real exposure likely extends beyond these direct dependencies because the package is commonly installed in backend services, developer environments, and CI/CD pipelines that integrate with Intercom's API. "The attack closely resembles the lightning@2.6.2 PyPI attack from earlier today, as well as the TeamPCP-linked supply chain campaign we reported yesterday affecting SAP CAP and Cloud MTA npm packages," Socket wrote. Neither Intercom nor Lightning immediately responded to The Register's requests for comment. We will update this story when we hear back from any of the compromised organizations. ®

Mini Shai-Hulud caught spreading credential-stealing malware

The wave of supply chain attacks aimed at security and developer tools has washed up more victims, namely SAP and Intercom npm packages, plus the lightning PyPI package.…

Give a man a phishing kit and he might get lucky a couple of times; teach an AI to phish and it'll change the landscape, if KnowBe4's latest phishing trends report is accurate. The cybersecurity and phishing awareness outfit released the seventh edition of its Phishing Threat Trends report on Thursday, and it appears that the internet's legions of phishermen are turning to AI in more ways, and more often, than ever thanks to their widespread adoption of AI. Nearly 86 percent of phishing campaigns KnowBe4 threat researchers have picked up on in the past six months have involved some sort of use of AI, according to the report. That's a gradual, steady increase over the past two years, too. 80 percent of phishing campaigns made use of AI in 2024, and 84 percent did so last year, suggesting holdouts are increasingly adopting the tech to broaden their reach.  That number may be troubling enough, but it's how AI is being used that KnowBe4 points out is the biggest problem. Well-written, highly personalized AI-crafted phishing messages are bad enough, but AI is also automating the reconnaissance and info gathering phases of a campaign, speeding up the phishing process and giving attackers more time to shift to multiple attack vectors to better gain their victims' trust.  While the report doesn't compile vectors as a share of total phishing attacks, it does note that there has been a 49 percent increase in phishing attacks that involve calendar invites, and a 41 percent increase in attacks that involve Microsoft Teams messages impersonating coworkers like IT support employees in order to harvest credentials and the like. Savvy multi-vector phishing operations still often start with an email, and that's one of the big areas where AI is broadening phishing horizons, according to the report. Automated reconnaissance enables attackers to comb through masses of information, extract target data, and feed that into AI-generated email lures. Those polymorphic phishing campaigns take a base template, jazz it up and make it unique to each individual, and voilà, a phishing message that's far less likely to be noticed than the typical one that relies on misspellings and bad grammar to weed out those capable of critical thought.  The report's data suggests that emails are only the start of the modern phishing campaign, however, as those increases in calendar invites and malicious Teams messages are often the second stage in an attack.  As IT teams are one of the most common groups impersonated by phishing attacks, one can easily imagine a phishing email followed by a Teams message from someone claiming to be from the help desk and demanding you click on a link to reset your password, or read and sign a new policy via DocuSign, etc. Both methods ultimately deliver credentials or remote access to an attacker, giving them what they were after. According to Microsoft, phishing campaigns involving AI lures are 4.5 times more effective than human-crafted ones. Meanwhile, the FBI says US cybercrime losses hit a record $20.87 billion last year, with phishing the most common complaint and AI-related fraud accounting for about $893 million of that total. ®

KnowBe4 says 86% of phishing it tracked used AI, and inboxes are only the start

Give a man a phishing kit and he might get lucky a couple of times; teach an AI to phish and it'll change the landscape, if KnowBe4's latest phishing trends report is accurate.…

China's "hacker-for-hire ecosystem has gotten out of control," according to Brett Leatherman, assistant director of the FBI's cyber division. This ecosystem includes private technology companies operating at the behest of the PRC's intelligence agencies while allowing Beijing to maintain plausible deniability.  "Motivated by profit, this network of private companies and contractors in China cast a wide net to identify vulnerable computers, exploit those computers, and then identify information that it could sell directly or indirectly to the PRC government," Leatherman told reporters on Thursday. Or, if the Chinese government won't buy it, the hackers-for-hire "turn from cyber mercenaries into cyber dealers," selling access to the compromised systems and stolen data to third parties on the dark web. "This leads to a less secure environment that is ripe for further lawlessness," Leatherman said.  Xu Zewei's extradition and the criminal charges against him, however, should send a message to China's contractor ecosystem, he added: "The protection you assume from operating inside China does not extend the moment you cross a border." Xu, a Chinese national, was extradited from Italy to the United States over the weekend and charged with nine hacking-related crimes. Italian cops arrested Xu last July. According to American prosecutors, China's Ministry of State Security (MSS) and Shanghai State Security Bureau allegedly directed Xu to hack thousands of computers and steal sensitive information in a way that hid the Chinese government's involvement. This happened between February 2020 and June 2021, and some of the digital intrusions were part of the 2021 campaign in which Hafnium (now better known as Silk Typhoon) exploited zero-day bugs in Microsoft Exchange and compromised hundreds of thousands of servers worldwide, including 12,700 organizations in the US alone. Other intrusions targeted American universities and researchers working on COVID-19 vaccines, treatments, and testing during the height of the pandemic, prosecutors allege.  The indictment claims that at the time, Xu worked as a general manager at a company named Shanghai Powerock Network, which the feds previously linked to Hafnium/Silk Typhoon. "Among other things, Xu worked on taskings from the SSSB, supervised hacking activity of other Powerock personnel in support of such taskings, coordinated hacking activities with fellow hacker Zhang Yu, and reported the results of the hacking activities to the SSSB," according to the indictment [PDF]. The indictment also charges Zhang, a director at Shanghai Firetech Information Science and Technology Company who allegedly operated at the direction of the SSSB, along with two unnamed SSSB officers who directed the hacking operations. Court records show Xu is charged with conspiracy to cause damage to and obtain information by unauthorized access to protected computers, to commit wire fraud, and to commit aggravated identity theft, which carries a maximum penalty of five years in prison; conspiracy to commit wire fraud and two counts of wire fraud, each carrying a maximum penalty of 20 years; two counts of obtaining information by unauthorized access to protected computers, each carrying a maximum penalty of five years; two counts of intentional damage to a protected computer, each carrying a maximum penalty of 10 years; and one count of aggravated identity theft, which carries a mandatory consecutive two-year sentence. Zhang remains at large, according to the DoJ. ®

One alleged cyber contractor was extradited to the US over the weekend

China's "hacker-for-hire ecosystem has gotten out of control," according to Brett Leatherman, assistant director of the FBI's cyber division.…

If you use Gemini CLI, watch out: Google has patched a CVSS 10.0 vulnerability in its command-line AI tool and is warning anyone running it in headless mode, or through GitHub Actions, to review their workflows. The update to Gemini CLI and the run-gemini-cli GitHub Action, published last week but largely unnoticed until one of the two credited research teams published its writeup on Wednesday, fixes a critical - and apparently easy-to-abuse - flaw tied to over-permissive workspace trust settings. Per Google's advisory published to GitHub, the issue stems from how the headless mode of Gemini CLI (frequently used in CI/CD environments and increasingly by AI agents) handles workspace folder trust: It automatically assumes any of the workspace folders it's active in are trusted for the purpose of loading configuration files and environment variables.  We trust you can see the problem here.  Novee researcher Elad Meged discovered the vulnerability (independently of Pillar Security's Dan Lisichkin, who Google also credited for the find), he told us, while studying CI/CD supply chain attack vectors.  "This vulnerability had nothing to do with prompt injection or the model 'deciding' to act maliciously," Meged told The Register in an email. "It was an infrastructure-level issue, where attacker-controlled content was silently accepted as trusted configuration and executed before any sandbox was initialized." A CVE hasn't been issued for the issue yet, but Meged told us Google has confirmed to him that it is in the process of assigning one. Novee also scored a bug bounty for the find, but declined to disclose how much.  A necessary fix, but expect fallout "This is potentially risky in situations where Gemini CLI runs on untrusted folders in headless mode," Google explained. "If used with untrusted directory contents, this could lead to remote code execution via malicious environment variables in the local .gemini/ directory."  Interactive mode in Gemini CLI does not behave the same way: it requires users to explicitly trust a folder before workspace configuration files are loaded, and the update brings headless mode into line with that behavior. The mitigations shipped in Gemini CLI versions 0.39.1 and 0.40.0-preview.3, but here’s the catch: the run-gemini-cli GitHub Action defaults to the newest Gemini CLI release unless users pin a specific version. In other words, anyone using the GitHub Action as part of a workflow without specifying a CLI version may have some cleanup to do. "GitHub Actions and other automated pipelines that rely on the previous automatic trust behavior will fail to load workspace-specific settings until they are updated to use explicit trust mechanisms," Google said.  The update may also break workflows that relied on Gemini CLI’s --yolo mode, which previously bypassed fine-grained tool allowlists and automatically approved agent actions without prompting. "In previous versions, when Gemini CLI was configured to run in --yolo mode, it would ignore any fine grained tool allowlist," Google explained in the advisory. "In version 0.39.1, the Gemini CLI policy engine now evaluates tool allowlisting under --yolo mode … As a result, some workflows that previously depended on this behavior may fail silently unless tool allowlists are modified to fit the task."  Those who do specify a version, says Google, ought to make changes to allow the newest, safest version to run and be prepared to fix those workflows anyhow.  Damned if you do, damned if you don't, in other words, but the fix is necessary, as explained by the folks at Novee Security, one of the credited finders. Across every workflow Novee tested the vuln on, the company noted, the results were devastatingly the same. "Code execution on the host running the agent gave an unprivileged outsider access to whatever secrets, credentials, and source code the workflow could reach," the Novee team explained. "Enough for token theft, supply-chain pivots, and lateral movement into downstream systems." In short, take action as Google suggests, or avoid putting AI agents in sensitive environments until the risks are fully understood. ®