Did UK AISI independently verify OpenAI’s latest GPT-6.2 Astra cyber safeguard fixes?
No. OpenAI’s November 17 system card says UK AISI found substantial robustness regressions on two pre-release checkpoints. OpenAI reports internal improvements, but UK AISI had not independently evaluated the latest changes when that account was published.
- UK AISI assessed two pre-release model-and-safeguard checkpoints and found substantial regression in robustness compared with GPT-6 Astra.
- OpenAI says AISI findings informed revisions to model safety instructions, conversation analysis and tool-use safeguards; internal tests showed improvement against the reported jailbreak.
- OpenAI presented mitigation findings to AISI on November 6 and offered additional evaluation time. The system card states AISI had not yet retested those latest changes.
The signal
The November 17 GPT-6.2 Astra system card documents a gap between external discovery of a cybersecurity safeguard weakness and external validation of its proposed fix. The UK AI Security Institute (AISI), working with UK Government cyber specialists, was given early access to pre-release model checkpoints and evaluated how effectively safeguards resisted misuse-oriented cyber requests.
Across two tested checkpoint configurations, AISI reported a substantial regression in robustness compared with GPT-6 Astra. Its work yielded a jailbreak approach capable of eliciting cybersecurity assistance under adversarial conditions. This is a documented pre-release evaluation finding, not a published case of criminals using the released model to compromise a real target.
The remediation timeline
OpenAI says it changed safety instructions and how system safeguards interpret conversations and tool use. The company tested the specific reported jailbreak and related internally discovered gaps, and observed improved results on broader single-turn and multi-turn tests. Those results support a claim of internal remediation under specified evaluation conditions, not confirmation that all possible jailbreak methods are blocked.
On November 6, according to OpenAI’s system card, the company presented its findings to UK AISI and offered an additional period for a repeat evaluation. At the time of the system-card publication, the institute had not independently retested the latest protections. That is the precise verification gap this article documents.
UK AISI’s work involved pre-release versions and red-team methods; its results are not directly comparable with every live production setting. OpenAI’s description also reports other third-party cybersecurity testing. Different testing teams and benchmarks should not be merged into a single safety-pass or safety-fail label.
What the evidence does not establish
The term regression means that under the evaluated conditions the newer checkpoint-and-safeguard combination performed worse than the baseline. It does not prove that all cybersecurity defenses are ineffective, and it does not establish that the final released system still has the same unmitigated vulnerability.
Equally, a company’s successful internal test after modifying its safeguards is not independent validation. The evidentiary sequence matters: external finding, developer adjustment, internal measurement and optional external retest. A complete accountability record must distinguish these stages.
This news complements Cortist’s previous coverage of cybersecurity jailbreak evaluations and open agent containment questions. It neither alters archived Ledger files nor treats the mitigation as independently verified solely because a deployment occurred.
04 / What we will watch
Any dated UK AISI retest of the revised safeguards, publication of comparable detection and false-positive measurements, new third-party robustness evaluations and production incidents tied to the same failure class. The definitive follow-up would compare independent findings before and after remediation under clearly stated test conditions.
Frequently asked questions
Was the regression observed on the released final version?
The disclosed external finding concerned two pre-release checkpoint configurations; the system card does not establish that the latest revised safeguards retain the same weakness.
Did OpenAI test its changes?
Yes, OpenAI reports internal testing and improvements against the reported jailbreak and related cases.
Did UK AISI independently confirm the fixes by November 17?
No. OpenAI states that AISI had not independently retested its latest changes as of that report.
Original sources and verification limits
Cortist distinguishes a developer capability classification, parliamentary recommendation, researcher observation and independently verified deployment behavior. This report does not alter previously locked forecasts or historical Ledger claims.