GPT-6 Cyber rolled out to select Daybreak Red users after OpenAI said the model went out of bounds less than 1% of the time in testing, down from about 5% for GPT-5.6 Cyber.
48.5% on the ExploitGym HoneyPot benchmark put the new model roughly 16 points above GPT-5.6 Cyber and ahead of GPT-6 Astra’s 42.4%, supporting OpenAI’s claim that it finds vulnerabilities more effectively.
Critical status under OpenAI’s Preparedness Framework marks a step up from GPT-5.6 Cyber’s High rating, while new checks are meant to block high-risk misuse such as attacks on power plants and water systems.
OpenAI said GPT-6 Cyber helped uncover 81 previously unknown Firefox vulnerabilities and harden its own training infrastructure, though executives said evaluation awareness remains a key unresolved risk.