OpenAI introduces GPT-6 Cyber for vulnerability research

OpenAI has unveiled GPT-6 Cyber, a model trained to identify software vulnerabilities, create working exploits and propose patches. The company says it found 81 previously unknown severe vulnerabilities in Firefox; Mozilla accepted the reports and has released fixes. The model also reassessed 30 known Firefox issues as more serious than previously believed.

In tests on the Linux kernel, GPT-6 Cyber reproduced 1,724 vulnerabilities and produced working exploits for 948. OpenAI says 304 critical findings were sent to maintainers with suggested fixes. The campaign scanned about 5,034 security-relevant locations, narrowed them to 2,308 candidates and used isolated virtual machines for confirmation. The reported compute cost was $1,213,430, or $1,279 per exploit.

On OpenAI’s ExploitGym benchmark, the model scored 48.5%, compared with 32.1% for GPT-5.6 Cyber. OpenAI reports that it respected operational boundaries in 99.93% of tests without special safety instructions and 100% with them. The figures have not been independently tested. Access initially is limited to selected Daybreak Red customers through the API and Codex, with broader availability planned.