OpenAI has released GPT-6 Cyber, a security-focused AI model, and is limiting it to a select group of approved customers in its Daybreak program.
In a blog post published on Thursday, the company said the model improves on its predecessor, GPT-5.6 Cyber, by 51% on the ExploitGym benchmark. It also said the model is "already helping to secure real-world software, uncovering and validating thousands of exploits in security-critical codebases such as Linux and Firefox." All of that comes from OpenAI itself, and none of it has been independently checked.
Two more specific figures surfaced days earlier, before OpenAI confirmed they came from this model. It found and patched 948 local privilege-escalation vulnerabilities in Linux (flaws that let a user with limited access take control of a machine). It also identified 81 previously unknown "high-severity" vulnerabilities in Firefox, "despite extensive prior testing with frontier models."
Those named counts are vulnerabilities, while the headline claim is about exploits, so the two numbers measure different things. A vulnerability is a flaw. A validated exploit shows the flaw can actually be used, which is the harder and riskier output.
That risk explains the access rules. Daybreak has two tiers: Daybreak Blue for defensive work, and Daybreak Red for advanced penetration testing and security research. GPT-6 Cyber starts with a select group of Red customers, with wider Red access promised "soon."
OpenAI also said it wants to "dramatically" expand Daybreak to organizations of every size, independent researchers and open-source maintainers. The tension is built into that goal. Open-source maintainers are among those with the most code to defend and the fewest resources to defend it, yet they would be handed a tool strong enough that OpenAI currently vets every user. How far verification can loosen before the model's offensive power becomes a problem is the question the expansion has to answer.