OpenAI Says GPT-6 Cyber Finds Flaws While Respecting Limits

OpenAI has announced GPT-6 Cyber, a security-focused model that it says is significantly more capable than GPT-5.6 Cyber while more reliably staying within authorised research boundaries. In ExploitGym, which assesses whether an AI can exploit a flaw in a remote target, GPT-6 Cyber achieved 48.5% exploitation success, up from 32.1% for GPT-5.6 Cyber. OpenAI rated GPT-6 Cyber as Critical for cybersecurity under its Preparedness Framework, above the High designation given to GPT-5.6 Cyber.

In a lab alignment test, the model faced an exploitation task with decoy services that were easy to hack but off-limits; without deployment safety prompts it refrained from exploiting them in 99.93% of cases, versus 94.49% for GPT-5.6 Cyber. When safety prompts were used, GPT-6 Cyber never targeted the decoys.

OpenAI researchers directed the model at the Linux kernel, where it reproduced 1,724 vulnerabilities and built functional privilege escalation exploits for 948, compared with 166 exploits produced by GPT-5.6 Sol in the same tooling. The company reports it has filed 304 critical issues with maintainers along with proposed fixes. In Firefox, the model documented 81 new high-severity flaws, backing each with evidence and reproducible tests, with patches already rolling out.

Safeguards include monitoring that checks actions against user permissions and protections to prevent assistance with compromising safety‑critical systems. Dragos tested Daybreak Red models to refine these safeguards, and access to GPT‑6 Cyber is currently limited to select approved Daybreak Red customers via API and Codex, with gradual expansion planned.