OpenAI has introduced the specialized GPT-6 Cyber model

OpenAI has announced the release of GPT-6 Cyber, a specialized neural network model designed for cybersecurity tasks. The tool is focused on the automated detection of software vulnerabilities and the development of corresponding patches to address the discovered flaws.

A key feature of the model is its built-in control system, which restricts AI actions to the scope of an authorized perimeter. This allows GPT-6 Cyber to be used for legal penetration testing and code analysis without the risk of exceeding defined security boundaries. The model was trained on specific datasets related to exploits and infrastructure defense methods. The system's effectiveness is confirmed by test results: the model has already discovered 81 previously unknown vulnerabilities in the Firefox browser and developed 948 working exploits for the Linux kernel, significantly exceeding the performance of previous versions of the tool.

Currently, access to GPT-6 Cyber is open to a limited group of Daybreak Red program clients. The model is already being actively used by large cybersecurity teams for Linux kernel research, confirming its practical significance in real-world conditions in the corporate sector.

The development is aimed at accelerating security audit processes, allowing specialists to identify critical code errors faster. Unlike general-purpose models, this version is optimized for working with vulnerabilities, which reduces the likelihood of false positives when analyzing complex systems.

For security teams, the implementation of GPT-6 Cyber means the ability to automate routine stages of red teaming and patch management. Using a specialized AI agent allows for reducing the time from the discovery of a potential threat to the preparation of a working fix, which is critical for protecting corporate environments.