OpenAI launches GPT-6 Cyber, an AI model built for cybersecurity

OpenAI has released GPT-6 Cyber, an AI model built for cybersecurity. The company says it is much better at finding, checking, and fixing software bugs than the previous GPT-5.6 Cyber. Its score on the ExploitGym test rose from 32.1% to 48.5%, a 51% jump. This test checks if a model can use given remote flaws.

In real use, researchers used the model to recreate 1,724 Linux kernel bugs and built working privilege escalation exploits for 948 of them. It also found 81 unknown high-risk flaws in Firefox and sent Mozilla proof, tests, and fix proposals. Mozilla has started rolling out fixes.

OpenAI has rated the model's cyber risk as Critical, higher than the last version's High. For now, only approved Daybreak Red customers can use it, with safety steps like abuse monitoring, sandboxing, and human oversight. OpenAI plans to expand the Daybreak program to help more groups with security research.