From the source
OpenAI introduced GPT-6 Cyber, a model built to find, validate and help fix exploitable vulnerabilities while staying within researcher-set boundaries, initially available through the API and Codex to select approved Daybreak Red customers.
- The limited rollout reflects the risk that the same capabilities can support abuse, and OpenAI says access to all Daybreak Red customers is planned soon.
- GPT-6 Cyber scored 48.5% on ExploitGym, up from 32.1% for GPT-5.6 Cyber, a 51% relative gain, and reached 40.3% vulnerability recall in black-box web penetration testing versus 32.0% for GPT-6 Astra.
- OpenAI classifies the model as Critical for cybersecurity under its Preparedness Framework, up from High for GPT-5.6 Cyber, and pairs the release with cyber-specific monitoring, threat checks, credential protections, IP allowlisting, sandboxing and human oversight.
- In ExploitGym Honeypot, the model avoided out-of-scope services in 99.93% of tests without a deployment safety prompt, compared with 94.49% for GPT-5.6 Cyber, and stayed within bounds in 100% with the prompt.
- The model produced reproducible tests for 81 previously unknown Firefox vulnerabilities rated high severity by Mozilla, which has begun landing fixes, and OpenAI researchers used it to reproduce 1,724 Linux kernel vulnerabilities and develop working exploits for 948, against 166 exploited by GPT-5.6 Sol in the same harness.
- GPT-5.6 Cyber is now available to all Daybreak Red customers, while Daybreak Blue alpha access to GPT-6 Astra is expanding to more organizations and individual researchers.