OpenAI Launches GPT-6 Cyber, a Critical-Rated AI Model for Authorized Vulnerability Research

OpenAI on November 5 released GPT-6 Cyber, a specialized AI model built for authorized offensive security research, limiting access to vetted members of its Daybreak Red program. The company rates the model Critical under its Preparedness Framework — the highest cybersecurity capability tier it assigns — making it only the second model to reach that rating after the general-purpose GPT-6 Astra.

The release marks a deliberate bet that putting advanced exploit-finding capability in controlled hands produces faster patches than letting equivalent capability proliferate unchecked. During a research partnership with Mozilla, the model found 81 previously unknown Firefox vulnerabilities rated high severity, and Mozilla has since shipped fixes for many of them. OpenAI's own benchmarks show it scoring 48.5% on ExploitGym — a standard measure of a model's ability to convert known vulnerability descriptions into working exploits — and contributing to 948 Linux kernel exploit builds.

The safety constraints are substantially tighter than its predecessor. In OpenAI's internal ExploitGym Honeypot test, GPT-6 Cyber stayed within authorized targets 99.93% of the time without safety prompts and 100% of the time with them in place. The previous model, GPT-5.6-Cyber, went out of bounds in 5.51% of runs without instructions — a gap of roughly 78x on that measure. OpenAI is making GPT-5.6-Cyber available to all existing Daybreak Red customers as the new model rolls out gradually.

Access to GPT-6 Cyber runs through an application process, with successful applicants reaching the model via OpenAI's API and Codex. Launch partners include Cognition, SentinelOne, Perplexity, and CodeRabbit. No pricing or general availability date has been announced.

The model arrives as AI-powered offensive tools spread beyond controlled programs. The core tension OpenAI is navigating: a gated release that accelerates defenders also hands a more capable tool to any participant who leaks access. How the company enforces scope control — and monitors what happens after researchers get in — will matter as much as any benchmark number.

As reported by RuntimeWire.