🇹🇼 Chinese, 5 November 2026, AI-powered vulnerability discovery
OpenAI Releases GPT-6 Cyber Model for Enhanced Vulnerability Discovery and Exploitation
OpenAI's GPT-6 Cyber model discovers 948 exploitable vulnerabilities in Linux kernel and 81 previously unknown high-risk vulnerabilities in Firefox.
OpenAI has released a new network security model, GPT-6 Cyber, which is designed to enhance vulnerability discovery and exploitation capabilities. The model is based on GPT-6 Astra and has been fine-tuned to improve its ability to find vulnerabilities, exploit them, and perform web penetration testing. In tests, GPT-6 Cyber discovered 948 exploitable vulnerabilities in the Linux kernel and 81 previously unknown high-risk vulnerabilities in Firefox.
The model’s performance was evaluated in various tests, including ExploitGym, Web penetration testing, and SRE-Bench. In these tests, GPT-6 Cyber demonstrated significant improvements over its predecessor, GPT-6 Astra, and other models. However, OpenAI notes that the model’s performance may be influenced by the quality of its training data and that it may not be able to exploit the latest vulnerabilities.
OpenAI has also demonstrated the model’s capabilities in a controlled red team exercise, where it successfully exploited vulnerabilities to escape a virtual machine and execute code on the underlying host machine. The model has also been used to identify vulnerabilities in the Brotli compression library and to generate patches for some of the identified vulnerabilities.