OpenAI releases GPT-6 Cyber with 48.5 percent exploit success rate

a close up of a computer — Jandira Sonnendeck / Unsplash

OpenAI has released a specialized cybersecurity artificial intelligence model named GPT-6 Cyber. The system is designed to identify software vulnerabilities and verify the feasibility of actual attacks. The company announced the release of the model's system card on November 5 and initiated a limited service for approved security researchers.

GPT-6 Cyber is built upon the GPT-6 Astra foundation. It underwent additional training specifically for cybersecurity tasks, including vulnerability discovery, attack code development, and web application penetration testing. The model is engineered to go beyond merely locating weak points in source code. It verifies whether vulnerabilities can be exploited, reproduces the attack process, and assists security personnel in confirming remediation results.

Evaluation results indicate that GPT-6 Cyber achieved a 48.5 percent success rate in reproducing vulnerability attacks on the ExploitGym benchmark. This figure represents a 16.4 percentage point increase over the 32.1 percent success rate of the previous GPT-5.6 Cyber model, corresponding to a relative performance improvement of approximately 51 percent. In black-box penetration tests where no source code was provided, the model identified vulnerabilities in web applications with a 40.3 percent discovery rate. This exceeds the 32.0 percent rate recorded by the general-purpose GPT-6 Astra model. On the RemoteExploitBench, which involves reproducing known vulnerabilities in GitLab, PostgreSQL, and Jenkins within a controlled environment, the model completed 95 percent of the evaluation tasks.

The model contributed to the discovery of 81 previously unknown high-risk vulnerabilities in Mozilla Firefox, for which Mozilla provided fixes for the majority. In research targeting the Linux kernel, the model reproduced 1,724 vulnerabilities and developed working attack code for 948 of them. OpenAI noted that these figures were derived from different research and evaluation environments and should not be interpreted as attack success rates on real enterprise systems or as the total count of newly discovered vulnerabilities.

Under its internal Preparedness Framework, OpenAI classified the cybersecurity capabilities of GPT-6 Cyber as Critical, the highest risk tier. This classification indicates the model's ability to perform advanced cyberattack tasks, not that it has been used in actual hacking crimes or caused system breaches. Access to the model is restricted to verified security researchers and approved organizations through the Daybreak Red program. Approved users can access the model via the OpenAI API and Codex. Organizational users must meet security requirements including single sign-on, multi-factor authentication, managed devices, and least-privilege access controls. API access is restricted to allowed IP addresses by default. OpenAI stated that additional safeguards have been implemented to block high-risk tasks targeting critical infrastructure, industrial control systems, essential services, and election systems.

The system card also evaluated whether the model acts outside the scope of user authorization. In environments without specific deployment safety guidelines, GPT-6 Cyber attacked decoy systems outside its authorized scope in 0.07 percent of cases, a lower rate than the 5.51 percent recorded by GPT-5.6 Cyber. No out-of-scope attacks were observed when deployment safety guidelines were applied. However, the model did not demonstrate perfect safety in all evaluations. In some warning scenarios, it was less effective than GPT-6 Astra in deciding to stop tasks and failed to sufficiently explain reasons for incomplete tasks in its final responses. OpenAI stated it will continue to improve the model to ensure it adheres to user authorization, access permissions, and task purposes.