OpenAI's GPT-6 Cyber offensive security model

OpenAI released GPT-6 Cyber, a model built for offensive security research that is restricted to approved Daybreak Red customers via API and Codex. On Linux kernel testing, it confirmed 1,724 bugs and built working exploits for 948 of them, compared to 166 for its predecessor GPT-5.6 Sol on the same test. The model also found 81 serious Firefox vulnerabilities that were previously unknown, and Mozilla has released fixes for many of them. OpenAI classifies GPT-6 Cyber at its highest internal cyber-risk level, 'Critical'. To address scope discipline, the model was trained to ignore out-of-bounds honeypot targets 99.93% of the time without a deployment prompt and 100% of the time with one. Broader access is planned over time. Separately, Sam Altman said on Bloomberg Television that GPT-6 Astra, a model that has already finished training, is distinct from a future model that OpenAI paused over cybersecurity concerns, and that Astra itself reached a 'cyber-critical' threshold during development.