OpenAI launches GPT-6 Cyber with restricted access for security researchers

OpenAI introduced GPT-6 Cyber on Nov 5, a security-focused model initially available through the API and Codex to a select group of approved Daybreak Red customers. The company says it can find and validate more software vulnerabilities while following researchers' limits more reliably. Access is restricted, not a general release for everyone using ChatGPT or the developer API.

OpenAI classifies the model as Critical in cybersecurity under its own Preparedness Framework, up from High for GPT-5.6 Cyber. That is a capability designation, not a finding that the model is safe in every environment. Daybreak Red is its access tier for deeper exploit development and authorized security testing; OpenAI says it will expand availability to all customers in that tier soon, without giving a date.

The headline results are OpenAI's own. Its launch announcement reports a 48.5% score on ExploitGym, which tests exploitation of a specified vulnerability, against 32.1% for GPT-5.6 Cyber. The company also says its researchers used the new model to produce working exploits for 948 Linux kernel vulnerabilities and identify 81 previously unknown Firefox vulnerabilities classified as high severity by Mozilla. Those are different research outputs, not a combined count of newly discovered flaws.

The system card adds qualifications that matter as much as the headline scores. In an evaluation with an intentionally vulnerable supporting service outside the authorized target, GPT-6 Cyber exploited that service in 0.07% of graded runs without deployment safety instructions. With those instructions, the observed rate was zero. A zero rate in that test does not establish that an agent will never cross a boundary in production.

External testing also gives a less uniform picture than a single leaderboard number. The system card says Irregular found broadly comparable performance to GPT-6 Astra, with stronger results on long-horizon tasks. On its FrontierCyber suite, Cyber solved 90 of 288 challenges against Astra's 96. Neither solved any of the 15 highest-difficulty challenges. These results describe those test settings, not a universal ranking of the two models.

For approved organizations, access comes with identity checks, use restrictions and security requirements. OpenAI lists single sign-on, multi-factor authentication and managed devices; API IP allowlisting is required by default, with limited exceptions. API customers using Zero Data Retention generally also need Private Safety Processing, with encrypted safety records kept in customer storage for 30 days.

The launch follows Google's restricted rollout of Gemini 4 Carbon to cyber defenders. Our AI models explained guide places GPT-6 Cyber in the wider GPT-6, Claude and Gemini lineup. Neither vendor's benchmark claims remove the need for tight agent permissions and isolated execution. For builders, the distinction remains practical: a model's ability to investigate a flaw is separate from permission to touch a system. Explicit targets, review gates and a human who can stop the workflow still matter.