OpenAI launches 'GPT-6 Cyber': its first 'Critical' risk rating, with exploit code for 948 Linux kernel flaws

While rumors said GPT-6.2 would land Friday, OpenAI's first release was a security model. On Nov. 5 (U.S. time) OpenAI announced GPT-6 Cyber. According to Yonhap Infomax it goes first to users of 'Daybreak Red', an approved offensive-security partner program, via API and Codex. It is the first OpenAI product rated 'Critical', the top cyber risk tier under its Preparedness Framework. OpenAI said it is a defensive tool for finding flaws before hackers do, but misuse could take down networks.

On the ExploitGym security test it scored 48.5% versus 32.1% for GPT-5.6 Cyber, and stayed within authorized scope more than 99% of the time. Researchers used it to find 81 high-severity Firefox flaws passed to Mozilla and to build exploits for 948 Linux kernel vulnerabilities (588 previously unknown). OpenAI announced that kernel research the day before. The top bug bounty for critical flaws in core products rose from $6,500 to $30,000.

OpenAI said security teams at Mozilla, Palo Alto Networks, Calif and Cisco already use it. A 'Daybreak Blue' track helps verified defenders validate vulnerabilities and fixes with fewer refusals, and access will widen to independent researchers and open-source maintainers. A 'Cyber at the Frontier' livestream is set for Nov. 12 at 10 a.m. PT.

Codex lead Tibo Sottiaux called it "in its own league on many cyber-specific use cases" (about 2,700 likes). A Calif user said GPT-6 Cyber is "out of this world" (about 540 likes). Asked when GPT-6.2 arrives, Sottiaux joked, "I live in the basement now." The Astra and Sol models did not appear.