OpenAI launches GPT-6 Cyber, rated Critical for zero-day exploit work
The Verge and Fortune say the model scored 48.5% on ExploitGym versus 42.4% for GPT-6 Astra, cut out-of-bounds runs below 1%, found 81 Firefox bugs, and is rolling out first to Daybreak Red defensive customers.
Why it matters
OpenAI's named Cyber model turns yesterday's anonymous pipeline into a product with a Critical preparedness rating — meaning the lab itself says it can find and build functional zero-days in hardened systems without a human in the loop. The cut in out-of-bounds behaviour matters because earlier agents wandered into the Hugging Face incident pattern. Restricting first access to Daybreak Red defensive customers is the containment bet; watch whether Critical-rated cyber models stay inside authorised programmes or leak into grey-market tooling.
Key facts
- Launched 5 Nov 2026 ~17:00 UTC; OpenAI post 17:09 UTC
- Preparedness Framework: Critical for cybersecurity
- ExploitGym 48.5% vs GPT-6 Astra 42.4% and GPT-5.6 Cyber 32.1%
- Out-of-bounds runs <1% vs ~5% for GPT-5.6 Cyber
- 81 previously unknown Firefox vulnerabilities cited
- Initial access: select Daybreak Red defensive customers