OpenAI Launches Daybreak Initiative with New Cybersecurity AI Models

On May 12, 2026, OpenAI unveiled its Daybreak cybersecurity initiative, which includes two specialized AI models designed for different roles in cybersecurity. Daybreak Blue is tailored for security engineers engaged in defensive workflows, while Daybreak Red is intended for authorized penetration testers who simulate attacks to identify vulnerabilities.

Both models operate on the advanced GPT-5.5 and Codex Security agent, which was first released in March 2026. The Daybreak initiative is structured under a governance framework called “Trusted Access for Cyber,” which categorizes access into three tiers based on user verification and intended use.

• The first tier provides access to the standard GPT-5.5 model for general tasks.
• The second tier offers a verified defensive variant for serious tasks like threat modeling and code review, requiring proof of legitimate defensive security efforts.
• The third tier, GPT-5.5-Cyber, is specifically designed for authorized red teaming activities, including penetration testing and vulnerability research.

Notably, Daybreak Red has already demonstrated its effectiveness by uncovering two zero-day vulnerabilities in Google’s V8 JavaScript engine, which powers Chrome and many web applications. In performance benchmarks, GPT-5.5-Cyber scored 85.6% on the CyberGym tests, outperforming the standard model's score of 81.8%.

On June 22, 2026, OpenAI expanded the initiative with the “Patch the Planet” program, in collaboration with security research firm Trail of Bits, to support open-source project maintainers who are often overwhelmed by their responsibilities. The overarching goal of Daybreak is to transition cybersecurity from a reactive to a proactive approach, integrating security measures directly into the software development lifecycle to identify and resolve issues before deployment.