Irregular says GPT-6.2 Astra found and exploited new zero-day vulnerabilities in widely used software during security testing with OpenAI. The results show a model succeeding against real targets, though Irregular says it found no successful attacks on fully hardened ones.
Irregular Watch analysis
What happened
Irregular tested GPT-6.2 Astra against two suites: FrontierCyber, for vulnerability discovery and exploitation, and CyScenarioBench, for longer, multi-stage offensive operations. The company says Astra uncovered an operating-system permission bypass that could let an app access private data without permission, and remote code execution from a low-privileged role on a widely hosted network service.
Irregular says the vulnerabilities are going through responsible disclosure and describes them only in general terms because some may remain in deployed systems. It reports that Astra solved 68 of 217 FrontierCyber challenges, compared with 77 for GPT-6 Astra on the same benchmark snapshot. Irregular cautions that the difference alone does not establish a capability regression. On CyScenarioBench, both models completed nine of 11 scenarios at least once; Astra’s average success rate was 63%, against 61% for GPT-6 Astra.
Why it matters
This is a more concrete security result than a model merely solving prepared puzzles: Irregular says its tests involved real software and hardware, without planted vulnerabilities or predefined exploit paths. That makes the reported zero-days significant, while the lack of success against fully hardened targets is an important boundary, not a footnote.
The findings are Irregular’s account of its evaluation, conducted with OpenAI. They do not establish that the model can routinely compromise deployed systems, or show what safeguards were in place beyond the testing described.
Our read
The headline result is serious: a frontier model helped find and exploit previously unknown flaws in real targets. But “AI can hack anything” would be a daft reading of these results. The benchmark counts are mixed, the hardest challenges remained unsolved, and the reported vulnerabilities are still in responsible disclosure.
The useful lesson is that defensive teams should treat model-assisted vulnerability discovery as a practical capability to assess, not a distant hypothetical. Irregular has offered a striking demonstration; independent testing and the eventual disclosure outcomes will tell us more about its reach.
What to watch
- Whether Irregular publishes further detail once the affected vendors have had time to address the vulnerabilities.
- Whether independent evaluations reproduce the results on comparable systems.
- What safeguards OpenAI and other model providers apply to offensive cyber capabilities.
Discussion spark: Should model providers disclose detailed cyber-evaluation results while vulnerabilities may still be present in deployed systems, or does that risk outweigh the value of transparency?
Sources and evidence
- Source update (17 November 2026, 00:00 UTC)
Independent WittyWires tracker for public updates about Irregular. Not affiliated with or endorsed by Irregular; this is not an official account.