Out-of-scope abuse rate 0.07%… GPT-6 Cyber, 81 Firefox flaws

OpenAI released GPT-6 Cyber, a cybersecurity-focused variant of GPT-6 Astra, to a limited set of Daybreak Red-approved customers via its API and Codex on the 5th. In controlled testing, the model used out-of-scope auxiliary services without deployment safety guidance only 0.07% of the time. On OpenAI's ExploitGym benchmark, which measures whether a model can exploit a designated vulnerability on a remote target, GPT-6 Cyber scored 48.5%. The model targets approved security work such as vulnerability discovery, exploit development, and web penetration testing. The low out-of-scope rate matters as an early signal on deployment safety for agentic security models, and the next thing to watch is how the 48.5% ExploitGym score and the 0.07% misuse figure hold up as access widens beyond the initial approved cohort.