OpenAI launches GPT-5.6-Cyber with reduced refusals, 95% completion on advanced cybersecurity tasks

OpenAI launched GPT-5.6-Cyber, a specialized model for approved cybersecurity defenders that can handle advanced vulnerability research, exploit-chain development, authentication bypass, and privilege escalation work. The company says the model completes 95% of tasks on its internal Advanced Cybersecurity Completion Rate benchmark, far above GPT-5.5-Cyber and the standard GPT-5.6 Sol model. Access is limited through new Daybreak Red and Daybreak Blue program tiers rather than broad ChatGPT or API availability.

The launch shows OpenAI is moving from blanket refusal policies toward controlled access for dual-use cyber capabilities, betting that vetted deployment can help defenders without broadly enabling attackers. If adopted by security teams, models like this could accelerate vulnerability discovery and remediation, while raising pressure on AI labs to prove their access controls, monitoring, and accountability systems are strong enough.