On November 5, OpenAI introduced GPT-6 Cyber, a restricted cybersecurity model that it says helped find 81 serious Firefox flaws and develop 948 Linux exploits that can provide higher-level access.
OpenAI said the Firefox findings had not been known before. Mozilla has released fixes for many of them. In separate research on the Linux kernel, OpenAI reported that the system identified 2,308 possible flaws.
It reproduced 1,724 vulnerabilities and developed exploits for 948 of them. OpenAI also said the model helped create possible fixes. The company presented the tool as a way to help defenders find and repair weaknesses faster.
Access is limited to organizations approved for OpenAI’s Daybreak Red program, which requires identity checks and other security measures, rather than being open to the general public. OpenAI rates the model “Critical” for cybersecurity. It reported a 48.5% score on its ExploitGym benchmark, compared with 32.1% for GPT-5.6 Cyber.
The company said the model stayed within permitted limits in 99.93% of tests without an instruction to deploy it, and in all tests with such an instruction. The reported results come from OpenAI’s research. The model’s ability to develop exploits also highlights the risks of a tool that could be used for both defense and harm, which is why its release is restricted.