OpenAI unveils Daybreak Blue and Daybreak Red cybersecurity models

OpenAI launched its Daybreak cybersecurity initiative on May 12, 2026, rolling out two distinct AI models designed to sit on opposite sides of the security fence. Daybreak Blue is built for security engineers handling defensive workflows. Daybreak Red is built for authorized penetration testers trying to break things before the bad actors do. The initiative runs on GPT-5.5 and the Codex Security agent, which OpenAI first released back in March 2026. Three tiers, one framework Daybreak operates under what OpenAI calls “Trusted Access for Cyber,” a governance framework that gates model access into three tiers based on what you’re trying to do and how much you’ve been vetted. The first tier is standard GPT-5.5, available for general use including basic blue teaming tasks. The second tier is a verified defensive variant designed for more serious work like threat modeling and code review. You need to prove you’re doing legitimate defensive security before you get access. The third tier, GPT-5.5-Cyber, is a more permissive model tailored specifically for authorized red teaming, including penetration testing and vulnerability research. Zero-days found in the wild Daybreak Red has already ...