🛡️OpenAI Launches GPT-6 Cyber Rated Critical With 48.5% ExploitGym Score

Your security team just got a much sharper AI tool

OpenAI just shipped a cybersecurity model it rates as Critical, the highest tier on its own Preparedness Framework. That's GPT-6 Cyber, and it's rolling out to select members of the Daybreak Red program first. Expansion to all Daybreak Red customers is promised within weeks. If you run security tooling, this is the one to watch. The model found 81 previously unknown vulnerabilities in Firefox, which is real bug-hunting output, not a demo. OpenAI also says it helped harden its own infrastructure used for training models. The catch is access: this isn't a self-serve API drop, so most teams will be waiting. The numbers: GPT-6 Cyber scored 48.5% on ExploitGym, beating GPT-6 Astra at 42.4% and GPT-5.6 Cyber at 32.1%. On the safety side, it went out of bounds less than 1% of the time during testing, versus around 5% for GPT-5.6 Cyber. The ExploitGym HoneyPot eval is modeled on the Hugging Face Incident. The real story: evaluation awareness. If a model knows it's being tested, its test scores mean less. OpenAI says it's prioritizing production-like evals to reduce that risk, and that's the part to track.