What's this about?
OpenAI released GPT-5.6-Cyber, a model built specifically for "advanced, authorized cybersecurity work," and expanded its defensive program "Daybreak," arguing the window to defend systems is narrowing. Access is deliberately restricted to vetted defenders, with extra monitoring for higher-risk use. This follows last week's disclosure that OpenAI's upcoming Astra model reached a "Critical" cyber-capability rating.
What exactly is a "cyber model"?
It's an AI model specialized in the security domain — finding software vulnerabilities, analyzing malware, reasoning about exploits and patches. The dual-use problem is obvious: the same skill that patches a hole can find and weaponize one. So OpenAI's play is to release the *capability* but gate *who* can use it — approved defenders only — which is why so much of the announcement is about access controls and monitoring rather than benchmarks.
Why is this notable?
It marks cyber capability graduating from "scary eval result" (last week's Critical rating) to a shipping, commercial product in one week. Per No Priors, Microsoft also launched an AI cybersecurity offering the same day — meaning the major labs are now simultaneously racing to sell defender-gated cyber models, the commercial flip side of the safety warnings.