OpenAI has introduced GPT-6 Cyber, a powerful artificial intelligence model specifically engineered for cybersecurity research. This new iteration significantly advances beyond its predecessor, GPT-5.6 Cyber, in its ability to detect, exploit, and help remediate software vulnerabilities under controlled conditions.
The model's capabilities were highlighted on ExploitGym, where GPT-6 Cyber achieved a 48.5% success rate, a notable 51% improvement over GPT-5.6 Cyber's 32.1%. This demonstrates its enhanced proficiency in developing working exploits for identified security weaknesses.
Detecting Critical Vulnerabilities in Major Software
GPT-6 Cyber is already making an impact in real-world security efforts. OpenAI reports that the model has been instrumental in identifying and validating thousands of software exploits across critical open-source projects. For instance, researchers utilized GPT-6 Cyber to uncover 81 previously unknown, high-severity security flaws in Mozilla Firefox, prompting Mozilla to release immediate software updates.
Similarly, the AI model assisted in developing working exploits for 948 Linux kernel vulnerabilities, with 588 of these being entirely new discoveries. Its robust performance extends to enterprise targets as well; in evaluations on RemoteExploitBench against platforms like GitLab, PostgreSQL, and Jenkins, GPT-6 Cyber successfully completed 95% of assigned tasks.
Enhanced Safety and Controlled Access
Given its advanced capabilities, OpenAI has elevated GPT-6 Cyber's cybersecurity risk rating from 'High' to 'Critical' within its internal safety framework. However, the company has also implemented upgraded safeguards to prevent misuse. These include continuous misalignment monitoring, IP allowlisting for high-risk tools like Daybreak Red, and targeted oversight specifically for critical infrastructure applications.
During black-box testing, where the model had no prior access to source code, GPT-6 Cyber achieved a 40.3% vulnerability recall rate, outperforming general-purpose models such as GPT-6 Astra. Furthermore, in ExploitGym Honeypot tests, the model stayed strictly within defined boundaries 99.93% of the time without requiring explicit deployment safety prompts, an improvement from 94.49% for GPT-5.6 Cyber.
Access to GPT-6 Cyber is currently being rolled out to a select group of cybersecurity partners, enterprise security teams, and trusted defenders. This controlled deployment occurs via OpenAI's Codex and API platforms, ensuring responsible integration into ongoing security research and defense operations.