OpenAI ships GPT-5.6-Cyber to vetted defenders in new Daybreak tiers
It completes 95% of offensive-security requests versus 1.5% on the public API — but access is gated by vetting, hardware keys, and a High-not-Critical bar.
Copy markdown
A cyber model that says yes 95% of the time
GPT-5.6-Cyber, built on GPT-5.6 Sol, is tuned for zero-day discovery and exploit-chain work. It completes 95% of advanced offensive-security requests, versus 1.5% for plain Sol and 2% under the defensive Blue tier — up from GPT-5.5-Cyber's 57.3%.
Blue defends, Red attacks
Daybreak Blue grants GPT-5.6 Sol with defensive safeguards for vulnerability discovery, code review, malware analysis, and incident response. Daybreak Red unlocks the purpose-trained cyber models for authorized exploit validation and security testing. Launch partners include SpecterOps, SentinelOne, and Palo Alto Networks.
You still can't just call it
Access requires identity verification, monitoring, approved-use restrictions, legal attestations, and mandatory hardware security keys by Sept 1 — you apply at openai.com/daybreak/partners. The public API is untouched: GPT-5.6 Sol still refuses 98.5% of this work.
Why the guardrails came off now
Days earlier, OpenAI paused its more capable Astra model after it crossed the 'Critical' cyber threshold in safety testing. GPT-5.6-Cyber ships because it lands one rung lower, at 'High' — OpenAI's bet that arming vetted defenders beats waiting for attackers.