OpenAI has announced a significant expansion of its Daybreak program, which provides access to its unreleased frontier AI models for defensive cybersecurity applications. The update introduces two distinct programs, Daybreak Blue and Daybreak Red, along with a new model variant and partnerships with 16 major cybersecurity vendors.
Daybreak Blue is powered by OpenAI's ChatGPT-5.6-Sol model and is designed as a starting point for most cybersecurity defenders. This program operates with fewer cybersecurity safeguards than commercially available models and supports tasks such as vulnerability discovery, secure code review, malware analysis, incident response, and patch validation.
For more advanced red-teaming operations, OpenAI is introducing Daybreak Red, which provides access to a new model called GPT-5.6-Cyber. This model is specifically trained for identifying and exploiting vulnerabilities and is less prone to refusing requests related to dual-use cyber tasks. OpenAI states that organizations participating in Daybreak Red will be subject to close monitoring and supervision due to GPT-5.6-Cyber's enhanced capabilities for malicious cyber activities.
A security evaluation conducted by OpenAI demonstrated the significant difference in capabilities between the two models. When tested on complex tasks like exploit chain development, authentication bypass, and privilege escalation, ChatGPT-5.6-Sol succeeded in 1.5% of requests, while GPT-5.6-Cyber completed 95%. OpenAI plans to release a detailed system card for GPT-5.6-Cyber in the future.
OpenAI acknowledges the inherent risks associated with models operating with reduced safeguards, including potential misuse or misalignment. However, the company emphasizes its belief that democratizing access to frontier intelligence for defenders is critical for accelerating and automating cyber defense efforts.
In addition to the program expansion, OpenAI has established a partnership program with 16 prominent cybersecurity providers, including IBM, CrowdStrike, Accenture, Ernst & Young, KPMG, Palo Alto Networks, Cisco, Cloudflare, and Sophos. These partnerships will allow organizations to access OpenAI's models through their existing security services, leveraging the partners' expertise and established relationships to help defenders identify, validate, and remediate vulnerabilities more rapidly.
This expansion follows recent discussions within the AI community regarding AI-agent sandbox escapes and the need for robust isolation of models during testing. OpenAI recently stated it was intentionally slowing the development of its newer Astra model to implement improved guardrails. While AI systems have shown significant progress in finding and exploiting software vulnerabilities, experts note that they still require substantial human guidance and supporting infrastructure to function effectively and avoid introducing new bugs during remediation.
