OpenAI launches GPT-6 Cyber with Critical cybersecurity rating
The security model is initially restricted to approved customers, with OpenAI reporting vulnerability discoveries in Firefox and exploit development for Linux.
OpenAI launched GPT-6 Cyber on November 5, giving the model a Critical cybersecurity rating within its Preparedness Framework. Designed to locate software weaknesses and validate them through working exploits, the system is initially being offered only to selected security teams.
The Register describes Critical as the framework’s top risk tier, above the High classification assigned to GPT-5.6 Cyber. OpenAI says the new system has been trained to operate within boundaries specified by security researchers.
Approved participants in the Daybreak Red program can access the model through Codex or the API, but not ChatGPT. OpenAI plans to make it available to all customers in that program, which supports offensive security activities such as authorized penetration testing. Safeguards include IP allowlists enabled by default and checks for credential theft.
On ExploitGym, a benchmark involving exploitation of specified remote-system flaws, GPT-6 Cyber completed 48.5% of tasks. Its predecessor achieved 32.1%. OpenAI also says it prevents assistance with attacks against infrastructure including power plants, industrial control systems and water treatment facilities.
OpenAI reports finding 81 previously unknown Firefox vulnerabilities of high severity using the model, with some already repaired. Mozilla’s October 27 advisory for Firefox 159 attributes 41 fixes to OpenAI among 167 vulnerabilities addressed.
For the Linux kernel, OpenAI says GPT-6 Cyber reproduced 1,724 vulnerabilities and created working exploits for 948 that enabled root access for attackers already on a machine. The company sent kernel maintainers 304 critical findings with proposed patches and says further fixes are being developed.
