OpenAI launches Daybreak — GPT-5.5-Cyber permissive model unlocks authorized red-team and pen-test workloads frontier safety policies had blocked
OpenAI unveiled Daybreak on May 10, a cybersecurity initiative built on GPT-5.5 plus a new sibling model called GPT-5.5-Cyber that is explicitly permissive for authorized red-team, pen-test, and vulnerability-research workloads. The launch acknowledges a structural problem in 2025-2026 frontier safety: blanket refusals on security-adjacent content blocked the legitimate defensive use cases as effectively as the malicious ones.
The permissive-tier framing is novel. Anthropic's approach has been single-model safety policies tuned for the lowest-trust user (which means refusing pen-test help to red teams that legitimately need it). OpenAI is now operating a two-tier policy: GPT-5.5 keeps the conservative defaults; GPT-5.5-Cyber is gated to authenticated security professionals and allows exploit-development, deep packet analysis, and reverse-engineering tasks that the base model refuses.
The bet is that authorization can be the safety boundary rather than capability — that the model can be allowed to generate things it currently refuses, as long as the requester is verified. This is the same architectural pattern as biological-research dual-use review: don't refuse the chemistry, verify the chemist. If it works, expect Anthropic and Google to follow with bounded-permissive tiers for biosecurity, weapons-research-adjacent, and other dual-use areas where the current refusal regime is starving legitimate defensive work.
VentureBeat — Anthropic vs OpenAI red teaming methods enterprise AI → · OpenAI — Findings from a pilot Anthropic-OpenAI alignment evaluation exercise → · VentureBeat — Four AI supply-chain attacks in 50 days →