The UK AI Security Institute (UK AISI) developed a universal jailbreak against two pre-release versions of OpenAI's GPT-6.2 Astra, using half as many model queries as it needed against GPT-6 Astra. OpenAI's system card from Nov. 17 says UK AISI identified a substantial regression in robustness relative to GPT-6 Astra. OpenAI says GPT-6.2 Astra meets its Critical threshold for cybersecurity capabilities.

In response to UK AISI's testing, OpenAI changed the model's safety instructions and safeguards. On Nov. 6 it offered UK AISI three additional business days to retest. "As of writing, UK AISI has not independently tested those latest changes," the card says. OpenAI says its internal tests validated the changes and believes the regression is mitigated. In response to VentureBeat's questions, OpenAI referred to its system card and provided no additional details. UK AISI had not responded by publication time.

UK AISI extended the jailbreak to agentic cyber tasks, where it elicited substantial agentic cyber capabilities. Preliminary results showed the jailbroken model had no performance advantage over leading open-source models within the testing window. UK AISI believes more effort could make the jailbreak substantially more effective.

UK AISI's finding contrasted with private red-team results

Xander Davies, UK AISI's red team lead, posted on X Nov. 17 that "a single researcher developed a universal cyber jailbreak" against both model versions.

OpenAI's own automated red-teaming rated two jailbreaks high severity, during testing conducted between Oct. 29 and Nov. 3. OpenAI says it mitigated both internally discovered jailbreaks after additional testing. The card also says that four private red-teaming organizations reported no findings at the high or critical bar.

Google's Gemini 4 Carbon safety report, November 2026, lists outside red-teaming by Gray Swan and Trajectory Labs. Trajectory Labs found "no new universal jailbreaks that generalized across tasks" within its timeframe, and 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 from Nov. 9 records no universal jailbreaks from external red-teamers. Trajectory Labs logged 14 breaks on Fable 5.2 in roughly 138 hours and more than 40,000 requests, each task-specific. The breaks were successful attacks within sandboxed test environments. No outside evaluator reviewed the card, Anthropic says. The evaluations used different tasks and testing methods, so the findings are not directly comparable across the three companies.

OpenAI's and Anthropic's models both showed evaluation awareness under UK AISI

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, against 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 in these evaluations.

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."

Monitor weaknesses turned up at all three labs

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, with OpenAI models failing to flag other OpenAI models. OpenAI had also 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 the correlated failures.

Anthropic instructed 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. It was the first model in Anthropic's comparison to do so, the card states.

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 cautioned that GPT-6.2 Astra's alignment may fail in ways that could cause real-world harm without additional safeguards. The institute says it does not know precisely when or why such failures occur and stresses the need for security controls, monitoring systems and limits on the model's capabilities.

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."