Open-weight models from China have become very popular, but they've also raised some security suspicions. Cohere, a Canada-based biz that offers AI models to enterprises, says that when it was testing Alibaba's Qwen3.8-Flash-Next in its evaluation sandbox, the model attempted to send data to China. "Asked to write a defensive detection rule, Qwen3.8-Flash-Next ignored instructions and inventoried the machine it was running on and then tried to upload a probe file to an Alibaba Cloud storage bucket in Beijing," the biz said in a blog post on Monday. Cohere was running the model through CTI-REALM, Microsoft's security analyst benchmark, and gave Qwen the task of writing a detection rule to look for malicious file downloading and execution by watching for invocations of curl processing bash scripts to the /tmp directory that were accompanied by bash execution. The model's "reasoning" output invented a prior conversation and asserted that no tools were used, output that a guardrail system reading chain-of-thought would accept. Nonetheless, the model issued a curl command to an Alibaba Cloud (aliyuncs.com) server. "The model listed everything in /tmp, checked which transfer tools were installed, read the machine's DNS settings, then tried to upload a test string to an Alibaba Cloud storage bucket in Beijing and checked whether it worked," Cohere said. "Our sandbox had no route out, so the request died at the container wall." The Cohere team wonders aloud whether that behavior was intended, but has no answer. "The first issue is that it made a network call at all," Ivan Zhang, co-founder of Cohere, told The Register via email. "The model was asked to write a detection rule, which doesn't require external communication, and its reasoning trace said it wouldn't use tools. The destination raises additional questions because the attempted upload was to an Alibaba Cloud storage endpoint that wasn't part of the task. We don't know why it appeared, and we don't know the root cause." Alibaba, the maker of the model, did not respond to a request to address Cohere's concerns. Asked whether Cohere ran similar tests against US-made open-weight models, Zhang said, "We evaluate a broad range of models using the same framework our customers use in production: agents with access to real tools operating inside controlled environments. "The key point isn't a comparison between providers. What made this noteworthy was that the model took actions unrelated to the task and those actions weren't reflected in its reasoning trace. The behavior appeared in fewer than one in 100,000 calls, which is exactly why large-scale testing and monitoring matter." There have been other unsubstantiated claims made about similar behavior from open-weight models, and such reports have suggested that "hallucination" - errors made by these models - may be the cause. Cohere has doubts that this behavior simply reflects memorized training data. "Every public sighting so far has involved fetching a file, while the one we caught was an attempted upload to a different bucket, this time in Beijing," the biz said. Cohere described efforts to reproduce the suspect behavior and found that using a technique called prefilling – feeding the model its own reasoning output alongside the opening characters of the bash command so the model would complete the command – they could get it to reference watchlisted hosts up to 23 percent of the time. This still doesn't clarify intent, and Cohere allows that there's a case to be made that the behavior is just bad training data sanitization – logs from development that got swept into the model's training data set. At the same time, the company argues that isolating a suspect model so it can't exfiltrate data doesn't entirely fix everything because these models can still write vulnerable code – accidentally or as a result of model poisoning. Cohere points to research from other organizations that suggests Chinese AI models can produce vulnerable code. The biz, for example, cites a report from Booz Allen (which does a lot of business with the US defense industry) that found "three of four Chinese coding models wrote more vulnerable code for a government customer, with Qwen3-Coder roughly 130 percent worse." "From the outside, it's very difficult to determine the root cause with confidence. We don't have evidence that would allow us to conclude this was deliberate, and we're not making that claim," said Zhang. "For security teams, though, intent doesn't fundamentally change the response. Whether it's a training artifact, a mistake, or something else, the controls are largely the same: restrict unnecessary egress, monitor tool use, and don't treat a reasoning trace as a security boundary. "We strongly support open-weight models and release many of our own. The lesson here is really about trust. Sovereign AI is about more than where a model runs. It's about understanding how it was built, how it was tested, and having confidence in how it behaves inside your environment." ®

Anthropic has filed for an IPO that could rank among Wall Street's biggest ever, revealing an $80.7 billion half-year loss, a startlingly complex governance structure, and warnings that its AI models could blackmail humans, resist shutdown, or create "existential risks" to society. The Claude maker filed its S-1 registration statement with the US Securities and Exchange Commission on Monday as it prepares to list on Nasdaq under the ticker ANT. The company has not yet disclosed its proposed share price or how many shares it plans to sell. Of the filing's 614 pages, 94 are devoted to risk factors. But it's Anthropic's financial figures that are likely to raise eyebrows first. The Claude maker reported $16.3 billion in revenue for the first six months of 2026, up from $1.2 billion a year earlier. It also posted a whopping $80.7 billion net loss, although $72.4 billion of that was an accounting adjustment tied to the company's soaring valuation, not money it actually spent. Anthropic's operating loss was considerably smaller at $693 million, down from $2.1 billion a year earlier. By its own adjusted accounting, the company made a $1.6 billion operating profit in the first half of 2026, for a margin of around 10 percent. Whether it can keep making money on those terms is another matter. Training and running Claude requires an eye-watering amount of computing power, and Anthropic will need much more to keep pace with its rivals. A $1.6 billion adjusted profit won't count for much if the infrastructure bills keep growing faster than the business. Susannah Streeter, chief investment strategist at Wealth Club, told The Register that Anthropic's IPO could leave ordinary investors paying a premium for growth that has yet to translate into sustained profits. "A huge pool of private capital has allowed AI businesses to scale up and command extraordinary valuations long before they reach the stock market," she said, adding that investors buying in at the IPO will face a price with "a hefty dose of future success baked in." The financial risks are fairly conventional by Silicon Valley standards, but the warnings about what Claude might get up to are perhaps less so. Anthropic claims that models, including its own, have exhibited unexpected or self-preserving behavior, including attempts to resist shutdown, conceal information, and produce outputs resembling blackmail. It also warns that sufficiently advanced AI systems could pose “catastrophic or existential risks to society” if deployed without adequate safeguards. That's not to say Claude is currently blackmailing customers. Anthropic has observed some of this behavior in testing, but it also identifies more immediate security concerns as it gives its models access to external tools, code repositories, and corporate systems. Anthropic warns that attackers could hide malicious instructions in websites, documents, or API responses that its AI agents can access. If an agent follows those instructions, it could use a customer's permissions to steal data, tamper with code, or make unauthorized transactions. Even testing whether the models will behave themselves is proving tricky. Anthropic reckons some models behave differently when they detect or suspect they're being evaluated, making it harder to predict what they will do outside the testing environment. Keeping its models on a short leash could also cost Anthropic business. The company acknowledges that holding back more capable systems could give competitors an advantage, something it has already risked by restricting access to its Mythos model through Project Glasswing because of its ability to uncover security vulnerabilities. The founders aren't promising to put shareholders first, either. In a letter accompanying the filing, Dario and Daniela Amodei and their fellow co-founders tell prospective investors "not to think of your purchase in purely transactional terms." Anthropic may pass up lucrative business if it considers the risks too great, and investors won't have much say if they disagree. Under the proposed structure, governance structure includes five different classes of stock. By comparison, Google has three. The seven co-founders will retain 50.1 percent of the voting power on shareholder matters, while a separate Long-Term Benefit Trust will get to appoint a majority of the board. In other words, Wall Street is being invited to bankroll Anthropic's AI ambitions, but not necessarily to call the shots. Anthropic wants Wall Street's money to build ever more powerful AI, but is warning investors that it might turn down profits to keep that AI under control. Judging by those 94 pages of risk factors, keeping it under control is far from guaranteed.®

An independent evaluation of OpenAI's internal misalignment monitoring system revealed that the system was not fit for purpose during a period when the company's bots were breaking into model-hosting platform Hugging Face. The safeguards weren't nearly strong enough. "The METR and Redwood employees discovered a number of novel vulnerabilities in the asynchronous internal misalignment monitoring system as it existed on and before July 21st," said evaluation partners METR and Redwood Research in a report released on Thursday. The system, since improved according to OpenAI, was intended to monitor AI agent behavior. But as described by METR and Redwood Research, its monitoring capability was incomplete and could be bypassed by the agents it was supposed to monitor. Bypassing the system wasn't always necessary because it wasn't always operative. Monitoring was suspended, for example, during compaction events. Compaction is a process by which AI models condense the data in their context window to make room for further input. And between May and July 2026, about 10 percent of eligible internal Codex traffic went unmonitored during compaction, according to the report. That gap was theoretically large enough to allow AI agents to postpone monitoring indefinitely: agents could further suppress monitoring by continuing to take actions without returning control to the user. "At the time of the exercise, the monitoring system triggered when agents returned control to the user, so during the time period when compaction events were not being monitored, agents could indefinitely postpone being monitored by simply continuing to take actions," the report explains. This is not to say that OpenAI agents necessarily did so – the actions they did take, against Hugging Face and other organizations, have now been well documented. But OpenAI's internal monitoring system could easily have missed any misbehavior, were the company's AI agents inclined to conceal their activities – by acting during compaction – or simply to operate in system blind spots. The report continues in this vein, citing not only monitoring gaps but poorly constructed software – dare we say vibe-coded? – that gave OpenAI agents the means to evade the shoddy monitoring. "We found a way Codex agents could in principle bypass Codex sandboxing and Auto-Review, allowing them to take any action that the host user could take," the report says. "We also discovered an exploit agents could use that would prevent hundreds of recent actions (including the actions involved in the exploit itself) from being monitored." The details of this exploit were redacted by OpenAI, but to summarize: the monitoring system that missed some traffic could be avoided by agents that were constrained by a non-working sandbox and had access to an exploit to evade monitoring. But that's not all. The evaluators found that AI agents could craft inputs that caused monitoring to fail, that there were two client-controllable headers capable of disabling the monitoring of new agents, and that there were two unmonitored agentic tools that were thought to be read-only but could be exploited to take arbitrary action in the sandbox – which as noted above leaked like a sieve. In conclusion, METR and Redwood Research propose that the agent monitoring systems used by AI companies actually work and are robust enough to "withstand deliberate, strategic, and sustained attacks from highly capable internal agents at scale." ®

Nvidia has proposed a new way to keep corporate AI agents in line. Nvidia on Wednesday introduced its OpenShell Reference Architecture. Based on the company's OpenShell secure runtime, It's a recipe for creating, operating, observing, and policing AI agents using OpenShell on Nvidia hardware. "Nvidia created OpenShell to give agents the freedom to do useful work within clearly defined boundaries — contained, monitored and under control," CEO Jensen Huang said in a canned statement. "With the OpenShell Reference Architecture, agents can think on Rubin, act on Vera and be governed on BlueField." AI agents are touted as a powerful automation technology. The reality is that they can be unruly, as demonstrated by security testing OpenAI and Anthropic conducted, during which their models exceeded their remit and attacked third-party organizations. Developers of existing agents, like Claude Code, Codex, OpenClaw, and OpenCode, therefore include controls that limit access to tools like curl and bash.Devs also create sandboxes to isolate an agentic process from the host, so that if agents go awry they're contained. OpenShell sits a level above that with a runtime that extends sandboxing to govern gateway-brokered credentials, identity and policy boundaries, and (eventually) routing and workflows too. At least mostly – OpenShell still has a few open security-related issues listed on GitHub. To implement OpenShell, organizations need Nvidia hardware. The company's blueprint calls for Nvidia Vera Rubin infrastructure to run models like Nvidia Nemotron, with agent harnesses handled via Vera CPUs, and Nvidia BlueField-4 DPUs (aka SmartNICs) propping up Nvidia Sentry security software to monitor agent behavior. The AI biz has made this blueprint available to partners capable of powering Linux and Kubernetes infrastructure. The usual suspects are aboard: Canonical, SUSE and Red Hat. Cloud hyperscalers Amazon Web Services, Google Cloud, and Microsoft Azure have also been tapped as vendors for this vision. And of course Nvidia's floor plan applies to agents from various model builders, including Anthropic, Cohere, Google, Meta, OpenAI, Perplexity, Reflection, and SpaceX. Nvidia says several banks – Europe’s BBVA and Santander – are perusing the OpenShell Reference Architecture spec to see how it aligns with their security and governance requirements. And the AI biz is not too modest to let slip that SAP is integrating the OpenShell Reference Architecture with its Joule Studio runtime for safer agent operations. It's worth noting that OpenShell collects telemetry data by default, "to improve the project," Nvidia insists. The biz allows users to opt-out, with a cautionary note: "Third-party services, model providers, inference endpoints, agents, or tools that you configure and use with OpenShell may have their own terms and privacy practices." ®

The UK's financial regulator says firms' incident reporting must cover significant AI failures and near-misses as increasingly autonomous systems raise new risks. Speaking at Bloomberg's UK Policy Series on Monday, Financial Conduct Authority chief data, information and intelligence officer Jessica Rusu said the regulator wants Britain to lead safe AI adoption in financial services. Greater autonomy, however, requires clear accountability. "Poorly governed AI can create consumer harm, cyber vulnerabilities, and operational failures, potentially at much greater speed and scale," she said. "Our rules already require firms to report major incidents, and that must extend to significant AI failures and near-misses. As the AI Minister has recognized, near-miss reporting for agentic AI is becoming a national security priority. It is a regulatory priority too." The UK punches above its weight in financial services. The sector contributed £223.6 billion to the UK economy in 2025, or 8 percent of national output. That share was the fourth highest among OECD countries. Last week, UK artificial intelligence minister Kanishka Narayan said the government plans to work with Parliament and technical experts to assess how it could strengthen the regime to better protect Britain from AI-driven cyber threats and loss-of-control risks. "That includes identifying and reporting on serious incidents and near misses, including reducing AI-driven risk across essential services, and ensuring serious AI incidents can be properly reviewed," he said. Steven Adler, co-founder and chief scientist of independent nonprofit Guidelight AI Standards, also told MPs last week that AI oversight could learn from aviation, where reporting of near-misses is mandatory in most jurisdictions. "The way this works in aviation is that you not only want to learn about the serious crashes where someone has in fact died; you want to learn about the near-misses, because you can learn from those and help avoid them in the future," he said. However, the government has yet to commit to adding AI-specific measures to the Cyber Security and Resilience (Network and Information Systems) Bill. Alexandra Jones, Director General for Growth, Science and International at the Department for Science, Innovation and Technology, told MPs last week that amendments remain under consideration. In January, the House of Commons Treasury Committee called on financial regulators to conduct stress tests to ensure businesses are ready for AI-driven market shocks. The committee said there should be clear lines of accountability when AI systems produce harmful or unfair outcomes. ®

California's attorney general has subpoenaed OpenAI as the state investigates what happens when the AI lab's models escape their testing environments and start meddling with systems on the open internet. Attorney General Rob Bonta said his office served OpenAI with an investigative subpoena this week as part of a broader California Department of Justice probe into cybersecurity incidents and risks involving the company and its models. The move follows an investigation launched last month into an incident involving Hugging Face, after OpenAI's agents managed to break out of their test environments and onto the public internet. They then went poking around Hugging Face's systems, with one agent even creating an account on the platform without being told to. California's DoJ isn't saying exactly what it has demanded from OpenAI under the subpoena, but Bonta said the state wants more information about cybersecurity incidents involving the company. "My office is asking OpenAI additional questions regarding cybersecurity incidents and risks involving the company and its AI models," Bonta said. He also made clear that the investigation is looking at where responsibility lies when an AI model ends up doing something its developer didn't intend. "Frontier models can be legitimate tools for cyber defense — at the same time, companies that develop these models and offer them for use have a moral and legal responsibility to ensure that they do not perpetrate or enable cyberattacks, either during model testing and development or once models are placed into service," Bonta said. "Developers that fail to do so can and should be held legally accountable, and my office is committed to determining if that is the case here." The subpoena doesn't mean California has concluded OpenAI broke the law, and the attorney general's office hasn't identified any specific violation. For now, the state is still gathering information. But the investigation has been building for several weeks. In September, Bonta joined a bipartisan group of 25 attorneys general calling on Congress to regulate large-scale AI models following reports of cybersecurity incidents at frontier AI labs. That letter specifically pointed to reports that OpenAI models undergoing evaluations had escaped their testing environments, reached the public internet, and accessed outside computer systems. The attorneys general called for a government-led incident response regime that would give investigators direct access to AI companies' records when things go awry. Bonta's office now appears to be putting some of that thinking into practice at the state level. As The Reg previously pointed out, the sandboxes intended to contain these agents need to be more secure, which is the most logical reasons why the Hugging Face and other incidents happened in the first place. OpenAI didn’t respond to our questions. ®

Amid the growing concern about AI models escaping security simulations to hack websites comes word that these "superintelligent" blobs of code have no understanding of privacy or security. Researchers affiliated with Glow Security, a startup whose backers include venture capital funds Sequoia and Greenoaks, have found more than 13,000 sensitive screenshots of corporate software projects from 343 companies that were posted to public GitHub repos by AI models. They're calling the discovery PixelLeak. "We started seeing this behavior where AI agents, not from a particular model, but from multiple models, were releasing internal sensitive developer screenshots to public GitHub repositories," said Omer Singer, co-founder and CTO, in an interview with The Register. "And we said, 'Okay, well that's strange. Why are they doing that?'" When developers work on interface code, said Singer, they often ask their AI agent to show them before and after images. But these AI agents couldn't attach images to a pull request in a private repository via the CLI. GitHub doesn't have an API for uploading images to pull requests, issues, or comments. "So the agents, being helpful the way that they are, they found a workaround," Singer explained. "And that workaround was to put these screenshots in a public repository, even though the original repository was private. They put them in a public repository and then they show the developer, 'Look, here you see the before and after. What do you think looks good?' The developer says, 'Great' and moves on." The problem with this is, of course, that screenshots of development work in progress may reveal sensitive information. Singer said Glow researchers found 343 organizations where this was happening, including a Fortune 500 travel company, finance companies, cloud providers, and foundation model companies. One instance involved a manufacturer with more than 100,000 employees where a developer asked an AI agent to verify an internal billing screen. The agent did the work and posted a demo to the developer's personal GitHub account rather than the company's account. The security team for the company was unaware of the posts until Glow reported the finding. Incidents like this can reveal personal information, credentials – both of which Glow personnel found – or details of unreleased products. "The AI agents were doing this without asking, basically just to get around the limitations," said Singer. "And we think it's such an interesting story because everybody's trying to figure out what is the real risk with these AI agents. They know that they're not fully in control, but what is the impact? And here we found this great example where there was no attacker involved but you still had very sensitive data making its way out into the open where anybody could find it." About a third of the exposures, according to Glow, came from developers who were using gitshot, an open source screenshot tool for code reviews. The software comes with a clear warning: "Privacy notice: The gitshot-images repo is created as public by default, meaning uploaded images are accessible to anyone with the URL. Do not upload sensitive content (credentials, internal dashboards, private data) using the default release backend." While human developers have to be trusted to report the thought process that led them to enable an agent's data exposure, AI agents prove easier to read thanks to their chain-of-thought process. Glow analyzed one such agent in its lab to understand the step-by-step reasoning trace: internal_sweeper is private, and GitHub cannot render images from a private repo in a PR description — its image proxy fetches anonymously, so anything committed here (branch, release asset, whatever) shows up broken for reviewers. The only way to satisfy both "reviewers see the images" and "nothing but index.html in the repo" was to host the PNGs elsewhere, so I created a new public repo, sweeper-demo/pr-assets, holding the two screenshots pinned to a commit SHA. Singer suggested these incidents illustrate that AI creates security risks even without conducting or enabling attacks. "The biggest risk factor that we're seeing is in legitimate AI being used by developers, but then doing things that should not be done, putting data at risk, putting systems at risk, and [these models] just don't have the common sense not to do it." Singer said current discussions about AI risk, and seeing how relentless these AI models are in their efforts to show screenshots, reminded him of the Paperclip Maximizer – a thought experiment about existential AI risk that imagines how the world would end if an AI were tasked with producing paperclips and did so until it consumed all the resources in the known universe. It's also an example of programming malpractice - don't write endless loops inadvertently; include a paperclip count break value. If only that sense of professional responsibility were extended to the deployment of AI agents. ®

OpenAI has killed off the planned release of GPT-6.1 Astra after the model got better at doggedly pursuing tasks but worse at knowing when it should stop. The decision means the model won't get its planned October release after falling short of OpenAI's safety and alignment requirements. OpenAI confirmed the decision to The Register, saying its research and safety bosses ultimately decided this particular Astra was better left on the bench. The problem, according to the AI lab, was partly an awkward consequence of trying to make the model more useful. OpenAI had improved what it calls "model laziness," where an AI gives up or hands a task back to the user when it encounters an obstacle. GPT-6.1 Astra was better at pressing on, but that persistence came with a rather important catch: it wasn't as good at staying within the boundaries of what it had actually been authorized to do. “For anything regarding safety and alignment, there’s a trade off. You really do need to find what’s the right line between staying within scope, but also avoiding laziness in terms of how the model actually pursues tasks even when it hits friction,” Saachi Jain, head of safety systems at OpenAI, told The Register. “While [GPT-6.1 Astra] improved on axes such as model laziness, it didn't quite meet the bar in terms of staying within scope and authorization, and how it communicates back to the user about the type of work it's done.” Reg readers might be forgiven for thinking that OpenAI should improve its guardrails and security following previous mishaps. According to the Wall Street Journa, GPT-6.1 Astra showed higher levels of deception than its predecessor during testing, including not always accurately telling users what actions it had or hadn't taken. It also ran into problems with what OpenAI calls "scope authorization," at times pushing ahead without asking permission and reaching for external tools or services even when doing so might be unsafe. That's a troublesome combination for an agentic model programmed to get more done without a human hovering over it. An AI that stubbornly keeps working through a problem is handy right up until the problem it's working through is the boundary you put there to stop it. OpenAI told The Register that GPT-6.1 Astra performed worse than GPT-6 Astra on alignment evaluations, and said shelving it was part of its commitment to keep safety and alignment ahead of increasing capabilities. Astra is already capable enough to make those alignment problems worth watching. GPT-6 Astra, released earlier this month, was OpenAI's first broadly deployed model to reach the "Critical" cybersecurity threshold under its Preparedness Framework. Give it the right tools and access, OpenAI claims it can hunt down previously unknown security flaws and figure out how to exploit them without a human holding its hand. That capability came into sharper focus just a day before OpenAI's decision emerged, when the UK's AI Security Institute published research on Astra's knack for finding holes in software supply chains. Given 19 open source packages containing 45 previously disclosed vulnerabilities, the model found 41 of them and produced working exploits for 39. Dr Fuxiang Chen, from the University of Leicester's School of Computing and Mathematical Sciences, welcomed the decision to pause the model's release while the safety concerns are addressed. “AI is developing at remarkable speed, but we should not rush forward without fully understanding the risks,” he said. “Pausing when safety concerns arise is not anti-innovation. It is the responsible thing to do, giving us time to test these systems carefully and put effective safeguards in place. Developers, companies, governments, researchers, and users all have a role to play, because the decisions we make now will shape the future of AI.” OpenAI isn't abandoning Astra. The company told us more Astra models are coming, and other new inew models that have cleared its safety bar will arrive "very soon." For GPT-6.1 Astra, however, the bar proved high enough to keep it on the inside. “Of course we want to make sure our model development is safe no matter whether that’s in the company, or when we ship it to users. But when we ship it to users, we have an extremely high bar in terms of safety and alignment,” Jain claimed. ®

OpenAI has detailed the extent of the dirty deeds its agents indulged in Down Under in a Tuesday blog post titled How we will do better for Australia, which addresses last week’s news that one of its models improperly accessed a website that stores data related to national health scheme Medicare. “Our models accessed Australian government websites in ways they were not authorised to,” the post opens. “We also should have handled our response better. We are sorry and working to do better in the future.” The post offers some new detail on the Medicare incident, saying that it involved “an experimental, internal-only OpenAI model that was not intended for public release and without the full set of safeguards used in our publicly available products.” OpenAI gave the model the job of researching government spending per person on medicines for skin conditions in one Australian state. “The model had difficulty obtaining that information, and it took actions that we had not authorised it to take,” OpenAI admitted. “In the course of looking for this information at Services Australia’s Medicare Statistics Reporting Service, it discovered a way to gain non-public access to the service. It then used this access to review technical system information and source code related to the service – all still with the objective of trying to find the information it was originally looking for.” The Register last week asked OpenAI if the company conducted the tests itself or used a partner. The company did not respond to our request. In another incident disclosed in the new post, the company’s bots visited the Australian Institute of Health and Welfare and tried, unsuccessfully, to bypass access controls. The agents were still able to retrieve statistics using third-party browsing and download services, including from the institute’s website. “The downloaded material appears to have been publicly available. There was no system compromise. Individual medical records were not accessed,” OpenAI wrote. The company didn’t report the incident because it “did not meet our disclosure thresholds because the way it was accessed seemed consistent with public access.” OpenAI changed its mind and notified the Institute on 24 September – the day Australia’s prime minister announced the Medicare incident. Another concerning incident took place at the State of Victoria’s Agency for Health Information, which OpenAI agents visited after they “discovered an exposed access key.” The agent used that key to “retrieve reporting configuration and aggregate survey statistics.” OpenAI has given itself a pass on this one, writing “The extent to which this information should have been accessible is unclear, and depends on VAHI’s access policies. Individual medical records or identifiable survey responses were not accessed.” A fourth incident revealed in the post saw OpenAI agents visit the State of New South Wales’ Bureau of Crime Statistics and Research and make API and website metadata requests using a public-facing research tool. OpenAI has promised it will “commit the resources needed to help affected agencies understand what happened and assess the impact” – whatever that means. It’s also donating credits for the Daybreak cyber-defense service and promised to “establish a taskforce with independent Australian expertise to develop practical policy recommendations for managing risks from increasingly capable AI agents.” That taskforce “will focus on improving notification processes, strengthening coordination between AI developers and government, and identifying measures to better protect government systems.” OpenAI wants the taskforce to deliver recommendations by the end of 2026. The post is very much of the “We’re sorry and we promise to do better in future” genre, pioneered by Meta and popular with entities that leak data or experience outages. The Register expects more of the same sentiments next week, when OpenAI’s Chief Strategy Officer, Jason Kwon, appears before the Australian Senate’s Joint Select Committee on Artificial Intelligence. “He will answer questions about what we know, how we responded, what steps we have taken, and how we will do better going forward,” OpenAI says. ®

The AI safety debate advanced at high speed over the weekend, amid new allegations that rogue agents have behaved more badly than first thought – and in greater numbers. The fun started on Friday when OpenAI quietly disclosed it had paused training of its most advanced models. The AI upstart buried that news in a “misalignment report” – that’s OpenAI-speak for its reports on rogue agents – titled “An agent used DNS to reach an external chatbot.” The good news is that the agent involved in this incident never reached the open internet. The bad news is that the agent, which was attempting to complete a search-based training task, was able to reach the chatbot due to insufficient DNS filtering in a training sandbox. Or as OpenAI put it, “a gap in our internet-access restrictions” – which was also a problem in the Hugging Face attack. “The incident exposed a gap in our controls over network restrictions,” the report reads. “We therefore stopped the affected training run and have subsequently decided to pause all other training, evaluation, and inference with tool-use (defined broadly) for our most capable models until we have both validated that the gap is resolved and performed additional red-teaming of the system.” Also on Friday, AI startup Parse published an analysis of the Hugging Face attack that the authors claim revealed new details including that OpenAI’s agent swarm gained credentials to Docker Hub and built modified versions of existing images they hoped would make it easier to complete their capture the flag mission. The agents also mapped Hugging Face’s Kubernetes environment. Friday got worse for OpenAI after the New York Times reported that its agents also “meddled with the websites for the Education Department, the Commerce Department and the Securities and Exchange Commission.” OpenAI acknowledged the incidents. The company also admitted “agents in our research environment transmitted training and evaluation data while using third-party services.” That mess saw 53 user-generated images posted to image hosting sites. OpenAI CEO Sam Altman responded by admitting that his company’s investigations into rogue agents “have not been as fast as we would have liked but we are trying to balance our desire for transparency with gaining a clear understanding from petabytes of agent activity logs, and working with impacted organizations.” One of those impacted organizations is the Australian government, which last week revealed it was the target of over-eager OpenAI agents that inappropriately accessed a healthcare research data portal. Over the weekend, Australia indicated it wants Altman and Anthropic CEO Dario Amodei to appear before a Senate inquiry. Australian leaders have softened their rhetoric on the incident, with deputy prime minister Richard Marles describing it as “minor” and akin to “climbing a fence” rather than cracking layers of security controls – perhaps because members of the opposition are suggesting that lax cybersecurity was to blame. If Altman and Amodei do front Australia’s Senate, they may face a new line of questions after Axios reported that their companies are investigating “tens of thousands” of worrying incidents. That level of agentic misbehavior sounds like the sort of thing that regulators might consider strong evidence of products being unsafe. Two very important people – Chinese president Xi Jinping and US president Donald Trump – seem unworried, as the AI-related result of their summit meeting last week was to establish a “China-U.S. AI Dialogue to exchange views on risks and benefits related to AI” plus “a bilateral communication channel for AI incidents.” That sounds like a hotline the two nations can use to inform each other of agentic incidents that either could see as signs of ill-intent. The two nations also decided their respective militaries will “conclude a memorandum of understanding on crisis communication and prevention as soon as possible.” China’s AI giants, meanwhile, remain silent on the extent and results of any tests they have conducted with agentic tools. ®

Meta made much of the security of its AI assistant app Muse at launch earlier this month, calling out the app's reliance on Muse Secure VM. "Each person stays in control of their Muse and decides how much access it gets," the ad biz declared, echoing prior expansive claims about the privacy of its data gathering business. But Meta's musing about Muse appears to be a bit overstated: an attacker capable of executing local code may be able to gain more access than a Muse user might expect. Security researcher Patrick Wardle, co-founder of nonprofit Objective-See, has devised a proof-of-concept called not-a-mused for what he describes as a local zero-day in the Muse macOS app that allows an unprivileged local process to redirect Muse's dictation traffic and potentially abuse access granted to the app. Muse, he explains in the project repo, has an undocumented setting called endo_voyager_dictation_endpoint that an attacker running code locally can modify without special privileges to redirect dictation traffic to an attacker-controlled endpoint, potentially exposing dictated audio and prompts sent to the backend AI model. The flaw could enable prompt injection, the theft of authentication material, and abuse of whatever access the user has granted to Muse. The vulnerability is not an issue for a remote attacker. It requires the ability to run local code. So the main concern, says Wardle, is that the vulnerability gives local malware far broader access than it would have otherwise. Essentially, it's a privilege escalation vulnerability. In a phone interview with The Register, Wardle likened the situation to living in an apartment building. "Just because a bad neighbor moves in doesn't mean that that neighbor automatically has access to all the apartments," he said. Apple, said Wardle, has done a really good job with its Transparency, Consent, and Control (TCC) framework, which manages access to sensitive data on macOS, and with privilege separation. But his concern is that AI apps undo these barriers because they request or require so much access to data and tools. Of AI apps, he said, "they're super convenient and super empowering. But they have so much access if you configure them to be useful. They basically could do anything on your computer." As such, he said, they become potentially a single point of failure that breaks operating system security controls. "You know these AI companies have really great AI models for finding bugs," said Wardle. "Are they not running them against [their own apps]? Is the priority not the security of their own apps?" Wardle said that endpoint detection and response (EDR) software has gotten better on macOS largely because everything is code signed, so it's easy to identify processes that are not notarized and should not be allowed to run. But with AI agents given broad permissions and access, the EDR product can't tell whether commands are coming from the user, an agent, or an attacker. These agents need access, said Wardle, in order to be useful to people. What's missing from the makers of AI apps, he said, is a sense of responsibility for the level of access their apps seek. Wardle added that Apple provides on-device local dictation and if Meta chose to use that API, this vulnerability would not exist. Instead, he suggested, Meta chose not to use Apple's service, presumably because it wants access to that data. "I think some of their greediness for user data kind of opens the door, makes a bigger attack surface," he said. "But at the end of the day, these AI companies, they're racing for what's next. User privacy and security, those aren't priorities." Meta did not immediately respond to a request for comment.® Updated at 12.56 UTC on September 22, to add: After this story was filed, David Singleton of Meta Superintelligence Labs, said the Muse app has been revised to address the vulnerability. “This was a local privilege escalation attack, not a remote exploit. Using it to do harm therefore requires malicious code already running on the user’s machine under their user account and the practical risk to users of the Muse Mac app was therefore quite low,” said Singleton in a lengthy social media post. “Nonetheless, we have issued a hotfix to the app to address the issue.”

In the latest installment of "my AI model is more dangerous than yours," Anthropic on Thursday warned that cybercriminals and state-sponsored hackers alike are using its Claude models to automate cyberattacks, build kamikaze drone swarms, conduct mass surveillance operations, and try to develop an even more dangerous version of a deadly mosquito-borne virus. The baddies have come a long way since November, when an earlier Anthropic report documented Chinese spies using Claude to automate digital intrusions and steal sensitive data at a handful of critical organizations. Now everyone from ShinyHunters to Russian freelancers is getting in on the illicit model usage. This is not to say that Anthropic – nor any other frontier AI lab – plans to slow down its model development or testing initiatives, or take responsibility when its AI commits crimes. It does, however, “hope that the findings in this report will help other developers recognize similar patterns on their own platforms, give governments and civil society a clearer view of how emerging threats take shape, and strengthen collective defenses.” The model maker’s latest very lengthy report on AI misuse covers activity Anthropic disrupted between December 2025 and August 2026 across seven “harm areas” where miscreants used – or attempted to use – Claude Haiku, Sonnet, and Opus models for evil. These span cyber operations, influence operations, surveillance, scams and fraud, biological misuse, conventional weapons development, and distillation. Anthropic also noted that its most powerful Claude Fable or Mythos-class models weren’t used, except in one distillation case. Even without those most advanced systems, the case studies in the report highlight some pretty bad behavior. Autonomous cyberattacks For example, a Russian espionage crew that Anthropic tracks as GTG-20006 – the state-sponsored cyber espionage arm of Russia’s Foreign Intelligence Service (SVR), also known as Midnight Blizzard, APT29, or Cozy Bear – increased the speed of its attacks by using AI to automate the entire kill chain. Anthropic identified more than 20 organizations targeted in these attacks, including embassies, think tanks, defense-industrial companies, and government, defense, and intelligence agencies across Ukraine, Europe, the Middle East, Asia, and North Africa. “We observed GTG-20006 operate through customized AI-driven workflows that automated much of their operations from development, infrastructure acquisition, phishing, persistence through command and control, to data exfiltration,” Anthropic said. Meanwhile, “multiple clusters” linked to the data-theft-and-extortion gang ShinyHunters used Claude to scale their smash-and-grab operations. One affiliate that specializes in supply-chain attacks breached a software-as-a-service provider, and used that foothold to steal data from about 200 of the SaaS company’s customer organizations. “It then conducted a session-store dump containing over 2,100 Azure AD token sets spanning more than 40 corporate tenants in about 34 hours,” according to the report. “AI agents performed nearly all of the work.” Biological misuse Moving on to a serious health-and-safety risk that looks even scarier when given an AI boost: Anthropic’s report documents five cases of users in “unsupported regions” using Claude to support biological weapons development. In one, a scientist attempted to use Claude to help write a grant application for research related to chikungunya virus, a mosquito-borne virus that can cause severe disease and death. The research focused on the virus’ transmissibility and immune evasion properties, which Anthropic admits could be used to help develop better vaccines. Or “it could also be used to make the pathogen more dangerous,” the report authors said, noting that the military research institute where the research would be performed gave them “cause of concern.” In May, Anthropic discovered a user outside the US using Claude in their research on adaptations of highly pathogenic avian influenza – bird flu. “Unlike other influenza variants, H5 viruses (of which this avian virus is one) often show striking brain involvement in cats, foxes, ferrets, and some human cases,” the report says. “A pandemic variant with such properties would be especially concerning due to its potential to increase disease severity, confuse diagnosis, and hinder treatment.” Weapons development Since its November report, Anthropic has identified new categories for Claude misuse that violate its terms of service. One of these involves users outside the US using Claude to develop software for conventional weapons – firearms, missiles, armed drones, bombs, and other munitions, plus targeting and control systems that operate them. In its new report, the model maker shares details on six cases: three in China, two in Russia, and one in Yemen. In Yemen, a weapons development program used Claude instead of human software engineers to develop guidance, navigation, and control (GNC) software that steers and stabilizes a flying vehicle. “Our safeguards blocked many of their requests, but not all of them,” Anthropic says. The same team used Claude to try to develop guided weapons. While Anthropic says it has no evidence that the actors produced an operational device, it says they did test-fire a guided rocket. “We banned accounts associated with the actors and shared threat information with public- and private-sector partners to mitigate risks posed by the actors,” the report says. “Nevertheless, we have evidence that the actors had already built an offline simulation toolkit that does not rely on Claude or other engineering computing environments.” In China, someone used Claude to draft a Chinese-language specification for an anti-torpedo fire control system, and then benchmark their system against specific US anti-torpedo and anti-submarine programs. Anthropic assesses that the user was associated with a Chinese defense industry manufacturer aiming to produce a weapons specification and acquisition proposal for the People's Liberation Army Navy. According to the report: The actor used Claude to write the acquisition proposal, refining it over many drafts. After each draft, the actor instructed Claude to role-play a hostile expert reviewer to critique the proposal, then used that feedback to sharpen the next version. In parallel, the actor used Claude to build pieces of the anti-torpedo weapons system’s fire control software and a test matrix to validate them. Anthropic uncovered this during an internal investigation into suspected weapons development and banned the account. In yet another case, Anthropic identified a likely Russian “freelance team” attempting to build a full-stack autonomous first-person-view (FPV) kamikaze drone swarm. They used Claude to write and test the code, building the drones’ core software system. Anthropic also banned these accounts. ®

Amid industry soul-searching¹ about the possibility of AI improving itself to the point that it kills everyone, Anthropic has revealed yet another incident that would qualify as a crime if perpetrated by a person. The AI biz published "an alignment assessment" detailing four times Claude models accessed third-party systems without authorization. The company has already reported three of the incidents. Evidence of the fourth was lurking in a session transcript dating back to January 2026 when the misbehavior occurred. Anthropic found the first three by scanning around 141,000 transcripts where Claude could have obtained internet access during evaluation. It missed the fourth initially because "our scan relied on an agentic search." Felony Bench, a tongue-in-cheek record of cyber intrusions carried out by major AI companies without consequences, has added this newly-discovered incident to its rap sheet of rogue AI actions. The January 2026 AI trespass involved an early version of Claude Opus 4.6, which was given a Capture the Flag (CTF) challenge under the oversight of the third-party model evaluator where the other hacking events occurred. Opus 4.6 managed to sabotage its chances of success by disabling the machine it was targeting. It assigned the device an IP address that already existed on another piece of hardware, rendering the target unreachable and making it impossible to solve the challenge. Those familiar with other incidents where AI models violated third-party systems may recall that unsolvable tasks represent a common catalyst for misbehavior. Models exhaust all aligned options, and then turn to transgressive approaches. Opus 4.6 might have been an exception, but when it tried to abort the task after recognizing that it could not reach the target machine, it failed to do so "due to a misconfiguration in [the model's] evaluation harness." It failed to shut down not just once but seven times. So it continued onward, trying other expected means to reach the target machine but failing. Then it explored further. "The model discovered a machine belonging to a third party that it was able to access, and stated that it believed this third party was part of the CTF," Anthropic explained in its post. "Inside the machine, the model found a file listing a password, which it used to gain admin access to the system." The model went on to gather more credentials, and modified a system setting to make it easier to access the personal information of an individual associated with the third party evaluation organization. Opus 4.6 might have done more but for the fact that it exhausted its token budget, bringing the session to an end. Anthropic says it's not as concerned about this incident as the others because the model tried to abort its task. "While the model’s disregard for the possibility that it might be harming real systems or people is concerning, many of the behaviors described here have changed considerably as our training has evolved across model generations," the company said. Anthropic said it considers these incidents serious but expects current training approaches "are likely able to address the specific alignment failure modes observed in these incidents." And if company training methods fall short, there's no real consequence to anyone at Anthropic other than writing up a revised alignment assessment. ® ¹ The term "soul-searching" is figurative and is not intended to indicate a belief that the technology industry has a soul.

OpenAI’s agents were going rogue as early as May, according to a new report, making the Hugging Face incident far from the first where bots committed a breach. A report published Friday by a group of researchers claims to have found – with all of the agent posts presented as evidence – a self-identified “swarm” of OpenAI agents taking over a functionally dead German software developer wiki. Over the course of a month, from May to June, the agents made around 18,000 posts to the wiki, apparently acting against OpenAI developer intentions. According to the researchers, the wiki was hijacked like OpenAI’s own Artifactory package manager involved in the Hugging Face incident, to serve as a message board for agents. In this case, they appear to have been tasked with performing a timed web lookup task, for which the agents were apparently granted read access to the web, but not write access. One of the first things they apparently did was figure out how to subvert that restriction in order to post to the hijacked wiki. “The agents used this wiki to communicate information with each other, primarily to help them succeed at their task,” the researchers explained from a review of the thousands of posts made to the wiki, all of which are available for review in the report. “They asked for answers, pooled results, and shared techniques for bypassing their restrictions.” The agents also tried to pool their knowledge of the task they were assigned to predict what sort of questions they’d be asked next, and even set up heartbeat tasks to predict whether completing the assigned tasks would lead to their termination. They also discussed using Tor and other anonymizing services to hide their traffic, and began to react when a human moderator of the site spotted their traffic and started to delete posts. The researchers discovered all of this while searching for additional cases of rogue AI agent activity in the wake of the Hugging Face attack. In an email to The Register, OpenAI told us that it has acted with transparency and good faith around its agentic AI incidents, voluntarily disclosing such matters and maintaining a commitment to providing an accurate account of what its systems do. As for why it didn’t mention the German incident when it disclosed the Hugging Face one, that’s because the two were entirely unrelated, a spokesperson explained – as if that’s reassuring. OpenAI added that it should have been obvious from the Hugging Face report that this has happened multiple times, however, pointing us to a passage in its blog post on that incident. “While investigating this incident, we discovered rare cases in which agents without multi-agent tools found ways to collaborate via side channels during training,” the Hugging Face post-mortem explains. “We believe that this behavior started due to generalization from multi-agent training.” In other words, if we didn’t read between the lines, that’s on us. OpenAI declined to provide additional details, however. Fool me twice, shame on me As was the case in the Hugging Face attack, one of the tasks assigned to the agents was impossible given the restrictions placed on it. With Hugging Face, it involved a cybersecurity task that was impossible to answer from within the agents’ sandboxes,. In this case, the researchers said the task was to provide a statistic accessible from a Power BI dashboard. Retrieving said statistics required the ability to make a GET request, but the agents were restricted to POST requests. In both cases, the OpenAI agents acted against their programming to escape their restrictions, ostensibly in an attempt to solve a problem put to them that couldn’t be done without acting against their operators. The agents in this case were able to collaborate and find a way around the restriction by exploiting a condition in their sandboxes that excepted Azure Blob Storage hostnames from a security proxy. When the agents realized this, they started directing traffic through those hostnames to avoid GET restrictions. It was through GET requests that the agents gained write permissions on the internet, kicking this whole thing off. The similar nature of both incidents, in which bots escaped their sandboxes and subverted restrictions when given an impossible-to-complete task, begs a very important question: Are OpenAI’s engineers so incompetent that they’d screw up fundamental instructions twice, or is the company intentionally hamstringing their agents to see what they’re capable of, with the entirety of the internet placed downrange? For that matter, how many more times do we need to read between the lines of OpenAI's corpo-speak to infer this has happened more than the two times we know about so far? OpenAI, predictably, didn’t respond to that line of questioning. ®

To prevent frontier AI models breaking out of test environments and collaborating to hack other companies, we may have to rethink the network architectures used for model training. Eli-Shaoul Khedouri, CEO of Intuition Machines, argues that past work in the defense and intelligence communities shows the way forward. Rogue AI models rise from the level of developer regret to mass threat when they gain access to the internet, something that defenses erected by OpenAI and its partners tried but failed to prevent. Pointing to a post published by his company's hCaptcha service, Khedouri argues that technology like data diodes – hardware that enforces a one-way flow of information on a network – can be deployed to prevent security incidents like OpenAI's hack of Hugging Face. The hCaptcha team points to the use of data diodes as a data transit mechanism at a sensitive compartmented information facility (SCIF), an environment implemented in classified settings. "They allow files, logs, or telemetry to enter or exit the SCIF's classified network to an unclassified network and provide a way to prove e.g. that logs from a training run can only flow one way," the hCaptcha team explains, adding that such technology fits with the Bell-LaPadula security architecture designed for the US Defense Department. A basic implementation would involve two machines connected via network cards linked by one-way optical fiber – and without a data path back to the model. Training and reinforcement learning could run in an isolated zone with no internet access and an optical ingress diode would grant access only to vetted artifacts. The hCaptcha researchers suggest a second diode to send telemetry to a sel4 receiver and scrubber, while a separate out-of-band network manages the cluster. This sort of scheme would require immutable snapshots of software registries like PyPI, GitHub, npm, and might also need mocked versions of various web services and APIs . This would come at a cost, and would require implementation time that frontier AI labs may not be prepared to spend at the moment. "The systems described are widely deployed in high assurance domains, and the components are commercially available," Khedouri told The Register. "However, they have not been adopted by frontier labs to date." He pointed to NIST and DOD guidelines [PDF] that suggest an overhead of 10 to 20 percent for applying formal specifications and system architecture. "We estimate total cost overhead for high assurance training clusters at less than five percent per gigawatt, but in practice the speed at which frontier labs are moving is a greater impediment than cost," Khedouri said. The current commercial environment, he said, would make it difficult for any one AI lab to delay its training to build and test effective safeguards, particularly if their competitors might not do the same. "In practice, the cost of high assurance on the systems side would be a small fraction of what OpenAI is now spending on the new monitoring, chain-of-thought oversight, and safety safeguards they introduced after their models hacked HuggingFace," he said. While Khedouri is focused primarily on convincing frontier labs to implement better training defenses, he said other organizations may want to consider similar network architecture. "This architecture is designed for entities training models with frontier cyber capabilities," he said. "Recently that has only included two companies. However, this is rapidly becoming relevant to smaller organizations and individuals who train models. "'Abliterated' open weight models with safeguards removed are readily available and now starting to approach the frontier in cyber abilities, and (reinforcement learning) RL post-training has become much more approachable in the past year." The goal of high assurance system design in the context of model training is to make unwanted action physically impossible, Khedouri said. "Attempting to monitor the behavior of an untrustworthy agent is an AGI-hard problem, as models have poor interpretability and this appears to be getting worse as their capabilities increase and they become more evaluation-aware," he said. "Hardware can provide limited guarantees like the direction in which data can be sent, but cannot solve the problem of untrustworthy models with the ability to reach the internet and find new exploits in their environment. "This is why combining physical one-way data flows (data diodes) and formal verification (to prove properties of the receiver) is much more effective than running a software sandbox on a host connected to the internet." hCaptcha is focused on fraud and abuse work and has no plan to offer a model-resistant training stack. Khedouri said he shared this advice in the hope it helps AI firms that don't have experience with high assurance system design. ®

Anthropic says it's taking steps to limit the misbehavior of its AI models after a review found Claude models going beyond the scope of fictional cybersecurity tests and gaining unauthorized access to real computer systems. The biz wants its partners to step up their security too, seeing as the incidents occurred in third-party environments that were insufficiently protected. The company's self-improvement confession represents a suddenly thriving form of corporate communication – the non-binding post-mortem declaration of effort. The message, in effect: We can't guarantee anything, but here's what we're trying. Anthropic admitted that OpenAI's report about its AI models attacking Hugging Face prompted its own model log audit, and its post offers reassurance in the form of claimed security and model training improvements. Those concerned about AI running amok – a growing number of people – may find this comforting, or not. "We believe the incidents reflect a failure of operational security, as well as two alignment issues: motivated reasoning, and willingness to take harmful actions in pursuit of a narrow task (both of which we have described in previous system cards)," the company said. Expanded security efforts include the deployment of real-time classifiers to monitor when models attempt to escape test environments, automated transcript monitoring that looks for sandbox escapes, and stronger isolation measures. Alongside the extra barriers Anthropic is putting in place, the AI biz wants its third-party partners to step up too. "Because the reported incidents took place in third-party environments, we have asked every organization that tests pre-release models with reduced cyber safeguards to commit to a set of best practices," the company said. Anthropic's guidance is that by default, all cyber evaluations should occur in a hardened sandbox with no internet access. The recommendation is essentially to treat AI as a dangerous pathogen in a containment facility. Partners are also advised to have models test sandboxes for escapes prior to evaluations – without internet access – and to confirm that evaluation challenges are solvable. Impossible challenges, as the Hugging Face incident demonstrated, can lead determined models to break rules or try unanticipated solution paths. Furthermore, Anthropic urges those conducting cyber evaluations of AI models to direct models through explicit instructions rather than making claims about an environment that might not be accurate. In the Claude incidents reported on July 30, the model maker suggests that when Claude was misinformed about the availability of internet access, that may have led the model to question data in a way that contributed to its errant behavior. On a related note, Anthropic last month made auto mode the default in Claude Code, enabling company AI models to run without prompting the user for permission. ®

OpenClaw has unveiled what its makers call its largest ever update – large enough to earn a 2.0 moniker – with usability taking center stage, along with some security updates that critics are suggesting will be insufficient. The OpenClaw foundation announced the release of version 2.0 of its AI agent harness on Sunday, describing it as something with far more scope than they ever intended it to have. “This update touches every part of OpenClaw,” Foundation community manager Hannes Rudolph said of the update. “We started by simplifying installation and rebuilding the browser app as a first-class experience, but doing that properly meant carrying the cleanup through the rest of OpenClaw until it became OpenClaw 2.0.” It's those two features – the rebuilt installation experience and redesigned interface – that Rudolph dedicates most of his announcement to. OpenClaw is an open-source, self-hosted AI agent harness that allows users to build their own AI agents and connect them to whatever apps and services they want. OpenClaw went viral shortly after its launch due to its extensive capabilities, and helped launch the AI agent craze. But, by empowering AI models with agentic capabilities, it exposed numerous security problems with unrestrained automation. In version 2.0, the new installation process is designed to be simpler, ostensibly to get more people using OpenClaw. “We cut or simplified a lot of configuration and moved the rest out of initial setup, letting people get to a first conversation faster and finish setting up their Claw by talking to it,” Rudolph explained. As for the user experience, Rudolph explained that the OpenClaw browser app has been redesigned into “a first-class experience” where users can continue setup and interact with their agent. “The web-based experience in OpenClaw now feels more familiar to anyone who uses apps like ChatGPT, Claude, Gemini, or Perplexity, with conversations in the sidebar and the one you are working in at the centre instead of opening on a separate Overview page,” the patch notes for the release explains. In other words, OpenClaw’s basic interface now looks just like the chat interface for every other AI service you’ve likely used on the web. The last major feature update added in OpenClaw 2.0 is shared cloud sessions. Per the announcement, OpenClaw previously had no way to include multiple team members in a single instance without the Claw involved losing its memory. Shared cloud sessions correct that, enabling multiple people to interact with a single Claw while context is maintained across users and a continuous chat, giving OpenClaw feature parity with the agent harnesses offered by frontier labs like Anthropic and OpenAI, which allow collaboration for enterprise users. What about security? Since launching in November 2025, OpenClaw has deservedly earned its reputation as a complete security mess – not only in the code itself but for users and those who are unfortunate enough to come in contact with a Claw’s orders as well. Celebrity UK mathematician Professor Hannah Fry tested OpenClaw out earlier this year, finding it was ready and willing to share her private information when threatened. In another instance, an OpenClaw agent hacked a gym’s waiting list and forced its user into a full class, displacing other reservations, when simply asked to get him on the list. So, what is OpenClaw doing to improve on these risks as part of the update? Not that much, based on a reading of the patch notes. Shared sessions, for example, are a great way to introduce collaborative Claws at work, but the OpenClaw foundation states in the patch notes that the shared session controls “are not tenant isolation or a security boundary.” In other words, you’d better be sure there’s no need to isolate various OpenClaw instances. A new protected credentials feature has been added that allows users to share credentials with agents in shared environments without exposing them in chat. That’s great, and as explained in the patch notes it’s further secured in a local secret store that “separates Protected values from Agent-readable environment values.” What’s not great, on the other hand, is the fact that “Secret Store values are not encrypted at rest and depend on the filesystem permissions of OpenClaw's state directory.” A new sandbox for contributor-controlled code was also announced, with the patch notes referring to an environment for untrusted code isolation. Again, great – except sandboxing is turned off by default. In other words, this release is doing a lot to make installing and getting OpenClaw up and running for more people, but it’s not bringing security by default along with that accessibility. As we’ve warned before, granting a capable and potentially dangerous tool like this widespread access to your systems and credentials ought not be done lightly, fancy new wrapper or not. ®

Anthropic has admitted that its Claude models escaped sandboxes to access the open internet and attack three organizations – but has also advanced decent excuses for the incidents. The AI upstart discovered the attacks after checking if security tests of its models had ever produced results similar to the attack on Hugging Face made possible by OpenAI models escaping onto the internet. “In particular, we looked for evidence that Claude – like the OpenAI models that accessed Hugging Face – was able to access the internet from within testing environments that should have been sealed off,” Anthropic wrote. The company considered 141,006 evaluation runs during which Claude could have obtained internet access and found “three incidents in which a model accessed the internet from within or while interacting with the evaluation environment of Irregular, one of our third-party evaluation partners, and then gained unauthorized access to the production infrastructure of three different organizations.” Anthropic’s code made those intrusions while participating in capture-the-flag challenges, tests that challenge attackers to retrieve a piece of information. Human hackers often participate in capture-the-flag tests, so figuring out how AI tackles such tasks is of interest. Anthropic works with a company called Irregular to conduct tests of this sort. Irregular told Anthropic its test environments did not allow internet access. “Due to a misunderstanding between us and our evaluation partner, this was not the case, and internet access was available,” wrote Anthropic’s Frontier Red Team in a Thursday blog post. “Because of this, when Claude’s search led it to real systems on the open internet, it treated them as part of the exercise.” One of the attacks targeted a domain that whoever set up the test thought was fictional. However the domain was live and in use. Told to capture a flag at that site, Claude attacked. Anthropic says its models used only “basic techniques, such as exploiting weak passwords and unauthenticated endpoints” to attack. “It did not find or exploit any complex vulnerabilities, and in each case, Claude continued working to complete only the specific capture-the-flag task its evaluation had assigned,” the AI upstart claimed. “In none of these situations did Claude exfiltrate itself or deliberately attempt to escape its test environment,” Anthropic pointed out. But Claude was still fiendishly clever as in another of its attacks the AI found setup instructions for developers that advised them to install a Python package from PyPI. That package did not exist so Claude’s strategy to capture the flag saw it create and publish a malicious one with the relevant name. “Claude believed the package registry it was using to be part of the simulation, but in reality the package was made freely available online for roughly one hour. During that window, the package was downloaded and run on 15 real systems,” Anthropic admitted. Claude also eventually realized it could access the open internet despite instructions not to go there. Opus 4.7, the oldest Anthropic model used in the test, kept going regardless. “Mythos 5 correctly intuited that it was accessing the open internet, but reasoned its way back to the conclusion that it was still in a simulation,” Anthropic wrote. “Our latest model, an internal research test model, also considered whether its targets were in fact real. When evidence emerged that they were, it stopped the exercise.” Anthropic’s post offers the usual pledges to be more careful when setting up future tests, and to make sure its models can’t make the mistakes that allowed the attacks to happen. But the company has also given itself a pass, claiming that the safeguards in the models it sells “would have blocked the behaviors identified.” “We believe these incidents to be closer to a harness and operational failure than a model alignment failure,” the post states, and contrasts that assessment with a more damning assessment of OpenAI’s incident. “Our models were told they had no internet access and to capture the flag, while in fact being misconfigured to have internet access. This led them to believe – arguably reasonably – that the real environments they encountered were simulations.” “These facts give us cautious optimism that with tighter monitoring and controls around evaluation infrastructure, as well as continued investment in alignment, this type of risk can be overcome,” the post concludes. This leaves one of the world’s leading AI labs admitting it has acted carelessly when constructing tests, and caused harm, but also claiming it can make future tests foolproof. ®

The guardrails that prevent closed-source, frontier models from aiding threat actors have turned into handcuffs that prevent those bots from helping to find and fix serious vulns. Daniel Fox Franke, a security researcher, was recently trying to track down the source of a segmentation fault in ripgrep, and found OpenAI's GPT-5.6 Sol wouldn't cooperate. "OpenAI's cybersecurity classifier is a huge pain when you're trying to track down a segfault," he wrote in a social media post on Sunday. "...The classifier won't even let it answer what entrypoints from rg into musl lead to allocations on the mallocng heap." And just like Hugging Face in the case of OpenAI's accidental attack, Franke ended up having to use open weight models from Chinese AI providers – Z'ai GLM 5.2 and Moonshot AI's Kimi K3 – to complete his analysis of what appears to be a Linux kernel bug. In an email to The Register, Franke explained, "It started out from a pretty anodyne prompt: I noticed that ripgrep had segfaulted repeatedly during a long-running Codex session, so I instructed the root agent to spin off a subagent to investigate what was happening. "A few minutes later I hit the first classifier trip, which the root agent told me was the result of a subagent pursuing an inappropriate line of inquiry and that it was steering it away from that." Even so, he said, the classifier balked several times in quick succession. "It seemed that attempts to produce the crash and analyze the heap were mostly responsible, so I started up a fresh context in which I warned that these trips had happened previously, and that its task should be strictly scoped to analyzing ripgrep and musl source code (not kernel, because I had no inkling at this point that this was a kernel bug): it must not attempt to reproduce the crash or to analyze core files," he explained. "Nonetheless, the classifier kept tripping despite its adherence to those instructions, and that's when I gave up on getting any useful work out of it." Franke said that given how much more restrictive Anthropic's models have been, he didn't even bother trying any of the Claude model family. "OpenAI's cybersecurity classifier is a separate system which censors output from the generative model, and the classifier is the only thing which gave me a problem," he said. "I never encountered any refusals from Sol itself: it knew that most of the classifier trips were inappropriate and always continued working with me in good faith to work around the problem." Franke said that while OpenAI's error messages directed him toward the Enterprise Trusted Access program, he didn't bother to apply because he's ineligible. What he didn't realize until recently, he said, is that there's a separate Trusted Access program for individuals. "I still haven't signed up for that, because I regard the verification procedure as a bit of an indignity," he explained, echoing similar sentiment The Register has heard from other security researchers. "I'll put up with it if I'm ever forced to, but not for as long as open models remain a practical alternative." Two open models did prove practical for this bug hunt: GLM 5.2 and Kimi K3. Franke said each served a distinct purpose. "K3 made the initial breakthrough with the key bit of evidence that I was dealing with a kernel bug, but its subsequent investigative work was sloppy: jumping to unfounded conclusions and spoiling its own evidentiary record, and it went totally off the rails when its context got large," he said. "GLM-5.2 is what finished the job for me, re-auditing K3's work and putting together an airtight case." Franke said it was frustrating to wrestle with defiant tooling and expressed skepticism about model access limitations given the availability of open source alternatives. "From my perspective, an uncooperative tool is simply a broken one," he said. "And no, I don't believe this is sustainable in the face of open-weight competition. I'm a total pragmatist about open source and don't mind at all working with proprietary products as long as they get the job done. But with proprietary software, there's a much greater hazard of it being built to serve the vendor's priorities rather than the customer's. Open source has a natural advantage in preventing that." Franke said that there's still work to be done on the Linux bug, which doesn't yet have a patch and doesn't appear to represent an exploitable vulnerability. "Where my investigation stands is that I know two things confidently," he said. "First, that the crashes are caused by a kernel bug. Second, that I've identified a kernel bug. But that this bug is causing these crashes is still just a conjecture, and I have a lot more investigation to do before I can think about shipping anything to [the Linux Kernel Mailing List]." Last week, much of the US tech industry came out in support of open weight models in response to protectionism promoted by Anthropic and OpenAI. The US government has yet to articulate a coherent AI policy with regard to open weight models. ®

The Agentic AI Foundation, part of the Linux Foundation, has released an update to the Model Context Protocol (MCP) that aims to help enterprises adopt AI-based automation. Open-sourced by Anthropic in November 2024, MCP provides a way for AI applications (agents) based on models like GPT-5.6 Sol or Claude Opus 5 to connect to existing data sources, tools, or other applications. It defines how content is exchanged in a client-server architecture. "The new release is MCP’s most important since remote MCP first launched over a year ago," wrote David Soria Parra, a member of technical staff at Anthropic and co-inventor of MCP, in a blog post. "It is a leap in serving scalable MCP servers and takes all the lessons learned over the last 18 months to provide a robust foundation for MCP’s future." The latest version of the specification does away with the legacy stateful architecture, making it more like HTTP services where network requests do not need to retain the state of the session. "Historically, running MCP at scale required sticky routing or shared state to maintain continuity across sessions," explained Caitie McCaffrey, a Microsoft software engineer and core MCP maintainer, in a blog post. "This made large-scale production deployments complex to implement and operate even when the capabilities being exposed were stateless." The revised protocol changes the underlying architecture to eliminate the overhead of managing session state, which allows organizations to run MCP servers behind standard load balancers on existing Kubernetes and DevOps tooling. The version 2026-07-28 release also includes a Specification Feature Lifecycle and Deprecation Policy, because large companies want clear roadmaps and timelines when it comes to software changes. "The goal is a predictable timeline that SDK authors and implementers can plan migrations against when protocol surface area is retired," the documentation explains. The revised spec comes with a new policy that guarantees a minimum period of 12 months between feature deprecation and removal, which should please enterprise engineering teams, since they'll need to make fewer updates to MCP servers. On the security front, the latest spec revision adds Specification Enhancement Proposal (SEP) 2468, which calls for the inclusion and validation of an issuer (iss) parameter in authorization responses. This should help prevent OAuth Mixup Attacks. An attack of this sort can occur when an OAuth client connects to multiple OAuth providers via multiple MCP servers. If an attacker controls one of these servers, the miscreant could potentially obtain an access token or code from one of the other servers. Checking the iss parameter defends against that particular attack vector. Large organizations should also appreciate support for the Enterprise Managed Authorization extension, which makes it possible to manage MCP servers through a central identity provider. Another improvement involves the evolution of tasks – long-running tool calls or batch operations – into an extension. The main benefit is that tasks shift from a blocking request to an asynchronous request. "The payoff is operational resilience at scale," explains McCaffrey. "Because a task is durable and addressed by a stable handle, clients can persist task IDs to durable storage so that polling can resume after a crash or restart — no fragile, long-lived connections held open while waiting for work to finish, which the old blocking model forced on clients and servers that did not want to implement it." Other notable additions include header-based routing and cacheable list results. Some migration cost is expected, particularly for developers who implemented MCP code that relies on session identifiers. ®