On November 5, 2026, OpenAI released a model it immediately classified at its own highest risk tier. GPT-6 Cyber is a security-specialized variant of GPT-6 Astra, purpose-built for authorized vulnerability research — and it has already been put to work finding previously unknown flaws in one of the internet's most widely used browsers.
Before its public launch, OpenAI used GPT-6 Cyber internally and found 81 previously unknown high-severity vulnerabilities in Firefox. That number is striking both for its size and for what it represents: an AI system autonomously discovering real, exploitable security flaws in production software trusted by hundreds of millions of people.
Critical by design
OpenAI rates GPT-6 Cyber at Critical under its Preparedness Framework — the highest category in its internal risk scale. That places it in the same tier as its predecessor model, GPT-6 Astra. For comparison, GPT-5.6-Cyber was rated only High. The company is not trying to downplay the risk; it is leading with it.
On its benchmark ExploitGym, GPT-6 Cyber completes 48.5% of tasks, compared to 32.1% for its predecessor — a substantial jump in raw exploitation capability. These benchmarks are self-reported by OpenAI and have not yet been independently verified.
Tight access controls
To use GPT-6 Cyber, researchers must go through OpenAI's Daybreak Red program — a vetted access tier requiring identity verification, additional account security steps, and agreement to strict usage policies. The model is currently available via the API and Codex, with a rollout to more Daybreak Red customers expected over the coming weeks.
OpenAI says the safeguards include more frequent model check-ins during high-risk operations, additional misalignment monitoring, and hard constraints that limit out-of-scope behavior. In OpenAI's own testing, the model violated its rules in only 0.07% of graded runs — but that is not zero, which is exactly why access remains restricted.
Who is already using it
Alongside its own internal research, OpenAI says security teams at Mozilla, Palo Alto Networks, and Cisco were among the early access partners. The GPT-6 Cyber announcement frames this as responsible vulnerability research — finding flaws before attackers do.
This kind of AI-assisted security tooling was once purely theoretical. The 81 Firefox flaws make it concrete. Whether this accelerates defenders faster than it accelerates attackers remains the central question — one OpenAI has not yet answered.