OpenAI Ships GPT-5.6-Cyber: Offensive AI for Vetted Defenders

OpenAI has introduced GPT-5.6-Cyber, a model trained specifically for offensive security research, and split its Daybreak program into two access tiers: “Daybreak Blue” for defensive work such as malware analysis and incident response, and “Daybreak Red” for exploit validation, pentesting and red-teaming, including access to the new model.

The key difference is the refusal rate: on OpenAI's advanced cybersecurity benchmark, GPT-5.6-Cyber reaches a 95 percent completion rate on sensitive queries — standard GPT-5.6 Sol sits at 1.5 percent, and the previous GPT-5.5-Cyber managed 57.3 percent. OpenAI argues the window between vulnerability disclosure and exploitation is collapsing as attackers adopt AI tooling.

The capabilities are demonstrably real: before launch, the model discovered two previously unknown flaws in Chrome's V8 JavaScript engine that could be chained to corrupt memory and bypass the V8 heap sandbox — registered as CVE-2026-15903 and since patched.

Access is deliberately hard to get: identity verification, legal declarations, ongoing monitoring, and mandatory hardware security keys from September 1. OpenAI positions the model explicitly as a tool to make defenders faster than attackers.

The pattern emerging across frontier labs is clear: build deliberately “more dangerous” specialist models, but gate them behind strict access and monitoring regimes. For companies, AI-assisted pentesting is about to become standard practice — the open question is who controls access.