GPT‑6 Cyber: Advancing the frontier of aligned cybersecurity agents
2026-11-06 · OpenAI
GPT‑6 Cyber: Advancing Aligned Cybersecurity Agents
What improved
OpenAI introduced GPT‑6 Cyber for discovering and validating vulnerabilities and developing fixes within authorized scope. It scores 48.5% on the official ExploitGym evaluation, compared with 32.1% for GPT‑5.6 Cyber.
Authorization boundaries
The model is trained to follow the boundaries researchers set. In the official evaluation, it avoided exploiting out-of-scope services in 99.93% of tests without the deployment prompt and 100% with it. Production use still requires explicit authorization, access controls, sandboxing, and human review.
Real-world use
Official examples include discovering previously unknown high-severity Firefox vulnerabilities and reproducing Linux kernel vulnerabilities with candidate patches. These are controlled security-research results, not a justification for unauthorized automated attacks.
Availability and adoption
Initial access is limited to approved Daybreak Red customers, with controls such as IP allowlisting and misuse monitoring. Confirm authorization, use isolated environments, and follow responsible disclosure. This article summarizes OpenAI's announcement and does not imply that the model is available on this platform.