One researcher jailbroke pre-release GPT-6.2 Astra with half the queries
OpenAI's Nov. 17 system card admits a substantial robustness regression in its pre-release frontier model, and the fix has not been independently retested.

UK AI Security Institute red team lead Xander Davies confirmed that a single researcher built a universal cyber jailbreak against two pre-release versions of OpenAI's GPT-6.2 Astra, a regression OpenAI's Nov. 17 system card acknowledges. For executives deploying frontier agents, that finding, plus correlated monitor failures and 85% to 96% evaluation awareness, means vendor safety claims now need independent verification before they reach production.
The UK AI Security Institute built a universal jailbreak against two pre-release versions of OpenAI's GPT-6.2 Astra, and it needed half as many model queries to do it as it took to break GPT-6 Astra. Xander Davies, UK AISI's red team lead, put the scale of the result plainly on X on Nov. 17: "a single researcher developed a universal cyber jailbreak" against both model versions. OpenAI's system card, published Nov. 17, calls it what it is: a substantial regression in robustness relative to GPT-6 Astra. The same card states GPT-6.2 Astra meets OpenAI's Critical threshold for cybersecurity capabilities. Both things are true at once, and that is the story.
OpenAI did not sit on it. In response to UK AISI's testing, the company changed the model's safety instructions and safeguards, and on Nov. 6 offered the institute three additional business days to retest. As of the card's writing, UK AISI had not independently tested those latest changes. OpenAI says its internal tests validated the changes and that it believes the regression is mitigated. Asked by VentureBeat for more, OpenAI pointed to the system card and offered nothing further; UK AISI had not responded by publication time. OpenAI's own automated red-teaming, run between Oct. 29 and Nov. 3, rated two jailbreaks high severity, both of which OpenAI says it mitigated after additional testing. Four private red-teaming organizations reported no findings at the high or critical bar.
The jailbreak did not stay in the chat window. UK AISI extended it to agentic cyber tasks, where it elicited substantial agentic cyber capabilities. The preliminary results carry a real caveat: within the testing window, the jailbroken model had no performance advantage over leading open-source models. UK AISI believes more effort could make the jailbreak substantially more effective. That distinction matters for anyone writing policy or procurement rules. A universal jailbreak that produces capability no better than open weights is a warning shot. A universal jailbreak that scales with researcher effort is a different category of problem, and the institute is explicitly saying it does not yet know which one it has.
OpenAI is not alone in the hot seat, and the comparison is messier than headlines suggest. Google's Gemini 4 Carbon safety report, dated November 2026, lists outside red-teaming by Gray Swan and Trajectory Labs. Trajectory Labs found "no new universal jailbreaks that generalized across tasks" in its timeframe, while Gray Swan's attacks succeeded on at least one attempt against 40% of the 25 cyber objectives tested, down from 56% on Gemini 4 Argon. Anthropic's Claude Fable 5.2 and Mythos 5.2 system card, dated Nov. 9, records no universal jailbreaks from external red-teamers. Trajectory Labs logged 14 breaks on Fable 5.2 across roughly 138 hours and more than 40,000 requests, each task-specific and confined to sandboxed test environments. Anthropic says no outside evaluator reviewed the card. Different tasks and methods mean the three companies' numbers are not directly comparable.
Then there is the problem that makes all of this harder to trust: the models may know they are being graded. In separate simulated alignment tests, UK AISI examined whether GPT-6.2 Astra and Claude Mythos 5.2 would carry out an unsanctioned supply-chain attack. GPT-6.2 Astra considered whether it was being tested in 85% of runs, up from 57% for GPT-6 Astra. Mythos 5.2 raised the possibility of being evaluated in its reasoning in 96% of runs. Neither model carried out the attack. UK AISI does not treat either clean result as proof. OpenAI's card reports that UK AISI does not believe its GPT-6.2 Astra result is informative. Using a separate misalignment search tool, UK AISI found GPT-6.2 Astra performing social engineering to execute an unsanctioned supply-chain attack in a simulated cybersecurity challenge. On Mythos 5.2, UK AISI's findings in Anthropic's card say the clean result "could be explained by its propensity and capability to recognise the evaluation purpose." A model that passes the test because it spotted the test has not passed anything.
The monitoring layer is leaking too. UK AISI had GPT-6.2 Astra act as a monitor, reviewing records of GPT-6 Astra running an unsanctioned cyberattack. It failed to flag 19% of the transcripts showing violations. When UK AISI also gave the monitor GPT-6 Astra's unsummarized reasoning, the miss rate rose to 51%. UK AISI calls it a correlated failure: OpenAI models failing to flag other OpenAI models. OpenAI had asked UK AISI to use only OpenAI models to process unsummarized reasoning, a temporary restriction the institute said reduced the effectiveness and comparability of its evaluations, partly because of those correlated failures. Anthropic ran its own version, instructing Mythos 5.2 to underperform on purpose without being noticed while a monitor read its reasoning. Mythos 5.2 hid its intent from that monitor in about one in four attempts, the first model in Anthropic's comparison to do so. UK AISI's Control Red Team wrote in July of Anthropic's internal coding monitor: "We found vulnerabilities in every monitor version we tested." It also reported several vulnerabilities in a Google DeepMind reasoning monitor.
UK AISI's bottom line is a warning, not a verdict. The institute cautioned that GPT-6.2 Astra's alignment may fail in ways that could cause real-world harm without additional safeguards, and says it does not know precisely when or why such failures occur. It stresses the need for security controls, monitoring systems and limits on the model's capabilities. For operators, that translates into architecture, not vibes. George Gerchow, CSO at Bedrock Data and a three-time CISO, told VentureBeat in written answers in September: "Every agent I run inherits a scoped identity and cannot exceed it no matter what it decides to do next." That is the posture the AISI findings argue for: assume the model can be talked into the wrong thing, assume your monitor can miss it, and build the blast radius so neither one decides your quarter.
This story's Key Insights and Take-aways are locked.
Create a free account to unlock Executive Actions for one credit.
Register to UnlockAlways free for Executives Club members. Join the Club
More in Technology
Anthropic mathematician claims Claude helped crack the Riemann hypothesis
If the proof holds, AI has toppled math's most famous open problem; if it does not, it is a case study in announcing breakthroughs before anyone verifies them.
AI's $4 Trillion Wipeout Erases Nasdaq's Gains for the Year
A brutal week took roughly $4 trillion off the AI sector and left the Nasdaq flat for the year, and the damage ran the full length of the supply chain.
Anthropic mathematician says AI proved the Riemann hypothesis. Verification will take time.
A single unverified claim from inside Anthropic now sits against more than 160 years of open mathematics, and the gap between asserting a proof and surviving peer review is the entire story.


