OpenAI on Monday launched GPT-5.6-Cyber and split its Daybreak program into Daybreak Blue and Daybreak Red, with the more permissive model available only to vetted partners. OpenAI says the model is tuned for authorized vulnerability research, exploit validation, and security testing, and its internal benchmark shows it completing 95% of advanced cyber prompts.
The point of the split is where the model sits: partners get the tool, customers get the findings. GPT-5.6-Cyber is built to answer higher-risk prompts around exploit chains, authentication bypass, and privilege escalation that a general model would refuse or heavily resist, which lets a security firm probe a client environment without handing that client the model itself.
For organizations that buy red teaming or vulnerability research, the control point is the vendor relationship and its engagement scope, not direct access to the AI. That concentrates frontier cyber capability inside approved firms, where the downstream exposure is whatever systems those firms are authorized to test.