OpenAI launches GPT-6 Cyber security model
OpenAI has released GPT-6 Cyber, the first model it rates "Critical" for cybersecurity, and says it is both much better at exploitation and much more reliable at staying within the scope researchers authorize.
OpenAI announced GPT-6 Cyber, a model trained to find and validate exploitable vulnerabilities. It is rated Critical under OpenAI's Preparedness Framework, up from High for GPT-5.6 Cyber. The release comes with benchmark and alignment numbers that go beyond the real-world results we covered earlier: 948 Linux privilege-escalation exploits and 81 previously unknown high-severity Firefox bugs, many of which Mozilla has since fixed.
On ExploitGym, which tests whether a model can exploit a specified vulnerability in a remote target, GPT-6 Cyber scores 48.5%, compared with 32.1% for its predecessor. On a new web penetration-testing evaluation, where models probe applications without access to source code, it found 40.3% of vulnerabilities, compared with 32.0% for the mainline GPT-6 Astra. On RemoteExploitBench, built from known flaws in GitLab, PostgreSQL and Jenkins, GPT-6 Cyber completed 95% of tasks. Astra running under Daybreak Blue refused all of them.
OpenAI says alignment was a central training focus. ExploitGym Honeypot, an evaluation informed by the Hugging Face incident, plants easy-to-exploit supporting services that are out of scope for the task. Without deployment safety prompts, GPT-6 Cyber stayed away from those services in 99.93% of cases, up from 94.49% for GPT-5.6 Cyber. With the prompts, it stayed in scope 100% of the time.
Access to the model is through Daybreak Red, OpenAI's tier for authorized exploit development and testing. Daybreak Blue gives verified defenders fewer refusals on mainline models. OpenAI says Cisco, Calif and CodeRabbit are already using GPT-6 Cyber. It also says it plans to expand Daybreak "dramatically" to smaller organizations, independent researchers and open-source maintainers. It did not give a timeline or pricing.