Overview
- OpenAI rolled out GPT‑6 Cyber to a small set of vetted Daybreak Red customers on Thursday and classified the model as "Critical" under its internal cybersecurity preparedness framework.
- In internal benchmarks OpenAI reported a 48.5% ExploitGym score for GPT‑6 Cyber versus 32.1% for GPT‑5.6 Cyber and said the model avoided out‑of‑scope services in 99.93% of tests without a safety prompt and in 100% of tests with the prompt.
- OpenAI said researchers used the model to produce evidence, reproducers and proposed patches for 81 previously unknown high‑severity Firefox flaws and to reproduce 1,724 Linux issues while developing 948 working exploits, and that it disclosed hundreds of findings to maintainers.
- The company added cyber‑specific controls including identity and account checks, more frequent model check‑ins for high‑risk requests, sandboxing informed by past escape incidents, and clarified that Daybreak Blue users can opt out of having session data used for training.
- Security teams and open‑source maintainers warn the volume of AI‑generated findings creates a heavy triage burden, raising questions about whether access controls and monitoring will be enough as OpenAI expands Daybreak and rivals offer similar vetted programs.