OpenAI has released GPT-6 Cyber, a restricted AI model designed to find and test software vulnerabilities. The company says the model uncovered hundreds of exploitable flaws in widely used software, highlighting the growing role of advanced AI in cyber defense and cyberattacks.
OpenAI's Critical Cyber Model
GPT-6 Cyber is specialized for authorized security research, including vulnerability discovery, exploit development, and web penetration testing. OpenAI classifies it as “Critical” for cybersecurity under its Preparedness Framework, a higher rating than the previous GPT-5.6 Cyber model. The company says GPT-6 Cyber scored 48.5% on the ExploitGym benchmark, compared with 32.1% for its predecessor (OpenAI).
Because of those capabilities, OpenAI limits access through its Daybreak Red program. The company requires identity checks and organizational safeguards such as single sign-on, multifactor authentication, and approved network addresses (OpenAI system card).
948 Linux Exploits Found
OpenAI says GPT-6 Cyber reproduced 1,724 Linux kernel vulnerabilities and developed working exploits for 948 of them. Researchers disclosed 304 critical findings to Linux maintainers with proposed patches. The model also identified 81 previously unknown, high-severity vulnerabilities in Mozilla Firefox, and Mozilla has begun releasing fixes for many of them.
Axios reported that human researchers reviewed the Linux findings before disclosure (Axios).
Restricted Access For Now
OpenAI says GPT-6 Cyber monitors activity outside a user’s authorized scope. Its safeguards also target high-risk requests involving industrial controls, essential services, and election systems. The company’s cybersecurity checks guide describes access controls and automated review for API users.
Security Teams Face Pressure
These capabilities give defenders a tool that can examine large codebases and validate weaknesses faster. They also increase pressure to fix flaws before attackers discover similar techniques. OpenAI plans to expand Daybreak access to more organizations, independent researchers, and open-source maintainers.
References
OpenAI. (2026, November 5). GPT-6 Cyber: Advancing the frontier of aligned cybersecurity agents. https://openai.com/index/introducing-gpt-6-cyber/
OpenAI. (2026, November 5). GPT-6 Cyber system card. OpenAI Deployment Safety Hub. https://deploymentsafety.openai.com/gpt-6-cyber
Axios. (2026, November 5). New OpenAI model targets 948 Linux security flaws. https://www.axios.com/2026/11/05/openai-linux-cyber-model
OpenAI. (2026, November 6). Cybersecurity checks. OpenAI API documentation. https://developers.openai.com/api/docs/guides/safety-checks/cybersecurity