OpenAI released GPT-6 Cyber on November 5, a model built specifically for vulnerability research and exploit development rather than general-purpose work. The company rates it Critical in cybersecurity under its Preparedness Framework — one step above the High rating it gave GPT-5.6 Cyber — and is for now shipping it only to a select group of approved Daybreak Red customers through its API and Codex, with wider access promised over the coming weeks.
The benchmark jumps are large. On ExploitGym, which tests whether a model can exploit a specified vulnerability in a remote target, GPT-6 Cyber scores 48.5 percent against 32.1 percent for its predecessor — a 51 percent relative gain. On RemoteExploitBench, a new OpenAI benchmark built on real GitLab, PostgreSQL and Jenkins deployments, GPT-6 Cyber completed 95 percent of tasks while the general-purpose GPT-6 Astra refused 100 percent of them. On an offensive web penetration test that hides source code, its vulnerability recall was 40.3 percent versus 32.0 percent for Astra.
The results are already touching shipping software. OpenAI says the model produced evidence and reproducible tests for 81 previously unknown vulnerabilities that Mozilla classified as high severity in Firefox, and showed that 30 known issues were worse than reported; Mozilla began landing fixes shortly after the reports. On the Linux kernel it reproduced 1,724 vulnerabilities and built working exploits for 948 — more than five times the 166 that GPT-5.6 Sol managed in the same harness — and 304 critical findings went to maintainers with candidate patches.
OpenAI's pitch is that capability and control can rise together. In a honeypot evaluation informed by the Hugging Face incident, where decoy services are easy to exploit but out of scope, GPT-6 Cyber stayed within bounds in 99.93 percent of runs without a deployment prompt and 100 percent with one, up from 94.49 percent for GPT-5.6 Cyber. Daybreak Red API keys require IP allowlisting by default, and the company has added cyber-specific misalignment monitoring.
Access is the contested part. Vetted defenders get the sharpest model; everyone else gets the general one. OpenAI is leaning on a $1 billion Daybreak for Frontline Defenders grant programme and a new Daybreak Red for Critical Infrastructure tier — refined with feedback from OT security firm Dragos — to widen the circle. Rivals are running the same play: Anthropic's Glasswing and Google's Fairwind programme both give approved defenders less-guarded frontier models. Gartner expects AI cybersecurity spending to reach $51.3 billion this year, nearly double 2025. OpenAI's own framing of the race is blunt: "we have a limited window to use frontier models to secure software."
