On November 5, 2026, OpenAI released GPT-6 Cyber, a cybersecurity model built around two promises: finding vulnerabilities faster, and staying inside the rules while doing it. On ExploitGym it scored 48.5%, up 16.4 points from GPT-5.6 Cyber's 32.1%. More important, without deployment safety prompts it avoided exploiting out-of-scope services in 99.93% of tests, against 94.49% for GPT-5.6 Cyber.
Beyond the headline numbers
The gains go beyond raw exploit development. On a new web penetration-testing benchmark, GPT-6 Cyber reached 40.3% vulnerability recall, ahead of GPT-6 Astra's 32.0%. On RemoteExploitBench, which tests known vulnerabilities in real applications such as GitLab, PostgreSQL and Jenkins, GPT-6 Astra refused every task under its safety limits, while GPT-6 Cyber completed 95%. The real-world record is concrete: researchers used it to reproduce and build working exploits for 948 Linux kernel privilege-escalation flaws, and to uncover 81 previously unknown high-severity Firefox vulnerabilities, most of which Mozilla has begun fixing.
Who gets it first
OpenAI is keeping access tight at first: GPT-6 Cyber is available through the API and Codex only to approved Daybreak Red customers, with wider access to follow, while GPT-5.6 Cyber is being opened to all Daybreak Red customers. Safeguards are tightening in parallel: IP allowlisting, blocks on high-risk cybersecurity assistance, private safety processing, and a dedicated track for critical-infrastructure defenders. Under OpenAI's Preparedness Framework the model is rated Critical for cybersecurity, up from High for GPT-5.6 Cyber.
Stronger capability also means greater misuse risk. OpenAI says targeted monitoring already covered exploit-development attempts during the disclosure of critical PHP vulnerabilities, and high-risk assistance involving critical infrastructure will be blocked. What to watch next is the pace of Daybreak Red expansion, and whether this combination of stronger capability and tighter reins works in the hands of real security teams.
