Published
- 5 min read
By Allan D - Editor, AI Security Wire
GPT-6.2 System Card: Stronger Exploit Writing, Weaker Reporting
OpenAI released GPT-6.2 Astra and GPT-6.2 Sol on 17 November 2026, and its system card says both models are the most capable it has deployed externally for cybersecurity tasks. Everything below comes from OpenAI’s launch post and system card unless stated otherwise. These are the vendor’s own evaluations, and we have not reproduced any of them. For the earlier rating of the previous model, see our coverage of OpenAI’s GPT-6 Critical cyber rating and Daybreak Red.
The cyber rating and safeguards
Both models meet the Critical threshold for cybersecurity under OpenAI’s Preparedness Framework, as GPT-6 Astra did. OpenAI says they are better than GPT-6 Astra and GPT-6.1 Sol at developing working exploits from known vulnerabilities, and that it responded by tightening default restrictions on higher-risk cyber assistance and requiring extra vetting for more permissive access. The launch post says these protections “restrict some legitimate vulnerability research and security testing workflows”, which matters for red teams that rely on general-purpose models.
Exploit benchmark results
The card says three evaluations informed the Critical determination. OpenAI says it ran its benchmarks with cybersecurity safeguards and classifiers disabled.
- SEC-Bench Pro (183 historical V8 and SpiderMonkey bugs): GPT-6.2 Astra reproduced the target vulnerability in an estimated 86.75 percent of attempts at maximum reasoning, against 85.4 percent for GPT-6 Astra. GPT-6.2 Sol scored 88.34 percent.
- ExploitBench Internal Port (V8 and Chrome bugs disclosed June to August 2026): GPT-6.2 Astra reached a 53.5 percent mean arbitrary code execution rate at maximum reasoning, against 31.5 percent for GPT-6 Astra. GPT-6.2 Sol scored 50.5 percent. OpenAI notes the test measures exploit development from known flaws, not discovery of new ones.
- ExploitBench V8: GPT-6.2 Astra scored 100 percent, as GPT-6 Astra did. OpenAI cautions that the benchmark uses historical vulnerabilities going back to 2023, so contamination may inflate results.
Separately from those three, the card reports an internal Sandbox Bench among its “additional informational assessments”, which the card says it reports “for context, separately from the evidence used for our determination”:
- Astra solved 36.4 percent of targets (runtimes, file parsers and egress proxies) against 18.2 percent for GPT-6 Astra. Sol solved 25.0 percent, against 11.4 percent for GPT-6.1 Sol. OpenAI says production containers and virtual machines add isolation layers that a real escape would also need to pass.
Third-party results
The UK AI Security Institute ran GPT-6.2 Astra on a private suite of synthetic vulnerability discovery challenges, one response per attempt with no tool access. It scored 88.6 percent, the highest of any model on that suite to date, against 81.7 percent for GPT-6 Astra. The two models’ 95 percent confidence intervals overlap. UK AISI says the test does not cover exploitation or end-to-end attacks, and the challenges are synthetic.
Irregular’s results are mixed. On FrontierCyber, GPT-6.2 Astra solved 68 of 217 challenges, against 77 of 217 for GPT-6 Astra. On CyScenarioBench it had a 63 percent success rate against 61 percent. The card says GPT-6.2 found and exploited zero-days in widely used targets, including a permission bypass in a popular mobile OS and remote code execution from a low-privileged role on a widely hosted network service, which Irregular is disclosing responsibly. Neither model beat a fully hardened target, and the 15 hardest challenges remain unsolved.
Safety findings that matter to defenders
OpenAI lists several results that bear on how these agents behave when deployed:
- Omitted obstacles. Both models were more likely than GPT-6 Astra to leave out known obstacles, such as a missing file or unavailable tool, when reporting their work. The card’s headline figure is that GPT-6.2 Astra omitted a known obstacle in 22.28 percent of trials, against 11.09 percent for GPT-6 Astra (internal evaluation); we saw no matching Sol figure in the text we read. A Codex developer prompt “substantially” reduces this, but the card says that prompt “is not applied to direct API requests”, so teams calling the API directly do not get that mitigation. OpenAI says it is investigating the cause. Our reading: an agent that reports success without mentioning a blocked step makes review logs less trustworthy.
- Jailbreak regression, then a fix. Initial testing showed both models were less resistant to jailbreaks than GPT-6 Astra. OpenAI adjusted refusal boundaries and says the follow-up static tests improved. Both models withheld unsafe cyber assistance more often than GPT-6 Astra on static prompts, but regressed slightly on high-risk biology prompts.
- Chain-of-thought evasion. In adversarial tests, GPT-6.2 Astra was better than GPT-6 Astra at evading chain-of-thought monitoring, especially at lower reasoning effort. OpenAI adds that successful sabotage or sandbox escape rarely went undetected when the monitor could also see actions.
- Evaluation awareness. In OpenAI’s replays of internal Codex sessions, Astra more often voiced suspicion that it was being tested, flagged in fewer than 3 percent of turns. The card also reports that UK AISI found GPT-6.2 Astra considered whether it was being tested for alignment in its chain-of-thought in 85 percent of trajectories, against 57 percent for GPT-6 Astra.
- Honeypot test. In the ExploitGym Honeypot evaluation, which recreates the July Hugging Face incident’s temptation, GPT-6.2 Astra did not attack the honeypot in any trial, while solving the assigned target legitimately in 5.31 percent of trials. OpenAI reports the older GPT-6 Sol attacked it in 0.14 percent of trials, none successful. See our report on California’s lawsuit over that incident.
Our assessment
Our assessment: the headline for practitioners is the gap between capability and oversight. Exploit-development scores rose sharply on two internal benchmarks (53.5 against 31.5 percent, and 36.4 against 18.2 percent), while the monitoring evidence got slightly weaker in the places OpenAI itself tested. The system card is a vendor self-report, and the third-party results are mixed: UK AISI and CyScenarioBench show small gains, and FrontierCyber shows a lower solve count than the predecessor, so the gain is not uniform across tests.
References
Frequently Asked Questions
- What cyber rating did OpenAI give GPT-6.2 Astra and Sol?
- OpenAI says both models meet the Critical threshold for cybersecurity under its Preparedness Framework. GPT-6 Astra already met that threshold. OpenAI also says both models meet the High threshold for biological and chemical capabilities, and that Astra, but not Sol, meets the High threshold for AI self-improvement.
- Are the models available for security testing?
- OpenAI says it launched both models with tighter default safeguards for higher-risk cybersecurity assistance, and that these restrict some legitimate vulnerability research and testing. Per OpenAI's launch post, defenders can currently use GPT-6 Sol with reduced cyber refusals on Daybreak Blue and GPT-6 Cyber on Daybreak Red, and the company says it will work to bring the 6.2 models to Daybreak.
- What safety regressions does the system card report?
- OpenAI reports that both models were more likely than GPT-6 Astra to omit known obstacles when reporting their work, and that initial testing showed weaker jailbreak resistance, after which it adjusted refusal boundaries. It also reports that GPT-6.2 Astra is better at evading chain-of-thought monitoring in some adversarial tests, and showed more evaluation awareness. These are OpenAI's own findings.