OpenAI Releases GPT-6 Cyber Only to Approved Red Team Customers

Builders evaluating offensive security tooling must first confirm Daybreak Red eligibility before any integration work. A conditional scenario: an internal red-team lead requests API access for an existing exploit-validation pipeline. The team would record the exact customer tier, the required deployment safety prompt setting, and the 99.93 percent scope-adherence figure from ExploitGym Honeypot. They would then test whether their current workflow triggers the 0.07 percent out-of-scope cases OpenAI declined to detail. Unknowns include whether the model will be offered outside Daybreak Red and how the Critical classification affects future rate limits or logging requirements.

Operators must verify Daybreak Red authorization and data-handling terms before any production use. A conditional scenario: a security program lead considers adding the model to an existing penetration-testing contract. The lead would document the approved customer list status, the phased rollout timeline, and OpenAI’s statement that the model reports results accurately. They would also note that OpenAI recently began denying internet access to frontier models during training. The evidence does not establish whether the same restriction applies at inference time for Daybreak Red customers or how retention policies differ from general API usage.