OpenAI Introduces GPT-6 Cyber

OpenAI has introduced GPT-6 Cyber, a specialised artificial intelligence model designed for authorised cybersecurity research. Announced on November 5, the model is reported to have helped researchers develop 948 Linux privilege-escalation exploits and identify 81 previously unknown high-severity Firefox vulnerabilities. According to OpenAI, it achieved 48.5% on the ExploitGym benchmark, compared with 32.1% for GPT-5.6 Cyber. The company also reported a 40.3% vulnerability recall rate in a web penetration-testing evaluation and successful completion of 95% of the tasks in RemoteExploitBench. However, the results varied across independent evaluations, and OpenAI acknowledged that performance depends on the target, task and testing conditions. Access to GPT-6 Cyber is initially restricted to selected Daybreak Red customers through Codex and the API.

OpenAI's Linux research involved examining 2,308 candidate findings, reproducing 1,724 vulnerabilities and developing 948 working local privilege-escalation exploits. The reported research campaign cost approximately $1.21 million in API credits. The company also highlighted Firefox security fixes credited to its researchers, while noting that these findings should not be treated as a comprehensive independent audit of the model. OpenAI has introduced safeguards to keep research activities within authorised boundaries, although its evaluations also identified areas requiring improvement in scope compliance and reporting. The company stresses that the model is intended for approved security research and that access does not permit testing systems without ownership or explicit authorisation. The development highlights the growing role of AI in vulnerability discovery, while underlining the importance of human verification, responsible disclosure and timely security fixes.