OpenAI extends 'Daybreak' security project and reveals new cyber model — but for approved users only

OpenAI has broadened the offerings of its Daybreak cybersecurity initiative, unveiling two new tiers: Daybreak Blue and Daybreak Red. Both tiers provide access to advanced models tailored for defensive cyber work, but with varying degrees of compliance. Daybreak Blue serves as the entry-level option, enabling users to engage in vulnerability discovery, secure code review, malware analysis, incident response, and patch validation.

Companies choosing this tier are granted access to cutting-edge models like GPT-5.6 Sol, which have been fine-tuned to comply with authorized defensive security tasks. On the other hand, Daybreak Red offers exclusive access to OpenAI's purpose-built cybersecurity models, specifically designed for authorized vulnerability research, exploit validation, and security testing.

This tier introduces the new GPT-5.6-Cyber model, built upon GPT-5.6 Sol and trained to excel in specialized cybersecurity tasks such as uncovering zero-day vulnerabilities and constructing exploit chains. Remarkably, this model demonstrates a higher level of compliance, successfully completing 95.0% of requests, compared to the previous 57.3% achieved with GPT-5.5-Cyber.

While OpenAI acknowledges the potential dangers associated with these models, they are only accessible through the Daybreak Cyber Partner Program, which currently restricts participation to approved organizations and individuals. The company emphasizes that using models with reduced safeguards poses significant risks, both from misuse and misalignment.

Despite these concerns, OpenAI believes that democratizing access to frontier intelligence for defenders is essential to accelerating and automating cyber defense. Interested parties can apply to join the program and leverage these powerful cybersecurity tools for their authorized work.