OpenAI is broadening access to its cybersecurity AI program with a new model built for authorized offensive work, warning that the technology is shrinking the time defenders have to react.
The program splits approved users into two access tiers. Daybreak Blue supplies frontier general-purpose models for approved defensive work; Daybreak Red supplies specialized cyber models for vulnerability research, exploit validation, and security testing. Only approved individuals and organizations get access, with identity verification and monitoring, and every individual Daybreak account must use hardware security keys starting September 1, 2026.
Acceptance rates show the gap between the tiers: in a vendor test of how often models accepted advanced cybersecurity requests instead of declining, GPT-5.6-Cyber took on 95% of them, while GPT-5.6 Sol answered only 1.5% and its predecessor GPT-5.5-Cyber came in at 57.3%. OpenAI’s Preparedness Framework rates the model “High” for cyber capability, one notch below its top “Critical” grade.
The model has also run against real software. During testing it turned up two bugs in Google’s V8 JavaScript engine that OpenAI says can be combined to corrupt memory and break out of V8’s heap sandbox. The company shared its findings with Google before going public, following coordinated disclosure practice.
Trained from the GPT-5.6 Sol lineage, the model prioritizes zero-day hunting and exploit chain construction while turning down fewer dual-use requests. The takeaway for defenders, OpenAI argues, is that the gap between a bug’s discovery and its exploitation will keep narrowing.
