OpenAI Releases GPT-6 Cyber to Hunt Software Vulnerabilities

OpenAI, the company behind ChatGPT, has released GPT-6 Cyber. It is a model built to find security holes in software and turn them into working exploits, attack code that actually breaks through a system's defenses. Researchers used it to find 81 previously unknown, serious vulnerabilities in Firefox, Mozilla's browser. In the Linux kernel, the core software that runs most servers, it produced working exploits for 948 vulnerabilities. 588 of those were previously unknown. That is more than five times the result of OpenAI's previous model, GPT-5.6 Sol, in a test run this past July. Access is currently limited to a small group of approved customers in Daybreak Red, OpenAI's program for advanced security work. OpenAI says wider access is coming in the next few weeks.

The same technique that finds bugs for defenders works just as well for attackers. That is why OpenAI is keeping access to the model limited. This is already OpenAI's fourth cybersecurity-focused model in 2026. Anthropic, the company behind Claude, and Google run similar vetted-access programs for security researchers. Controlled, limited releases are becoming the industry's standard answer to AI that finds exploits faster than people can fix them.