OpenAI has announced GPT-6 Cyber, an AI model built for security research. The company says it is far better at finding and fixing security flaws than GPT-5.6 Cyber, and its most notable result for everyday users involves Firefox.
During a research campaign, GPT-6 Cyber found 81 new flaws in the browser, all of which Mozilla rated as high severity. Several earlier AI models had already tested Firefox, so these bugs slipped past a lot of scrutiny. The model also proposed patches and showed that 30 bugs already on record were more dangerous than reported. Mozilla has since released fixes for many of them, and Firefox engineer Brian Grinstead says those fixes reached millions of users.
Patching these flaws early closes them off before attackers can find and use them. Firefox users who have automatic updates turned off should install the latest version manually.
Testing Linux and OpenAI’s own servers
GPT-6 Cyber also built working exploits for 948 Linux kernel vulnerabilities. OpenAI’s older GPT-5.6 Sol managed 166 using the same research setup. OpenAI has reported 304 critical findings to Linux maintainers along with suggested fixes, and companies like Cisco and CodeRabbit already use the model for code review.
OpenAI also tested it on its own infrastructure. In controlled tests, agents running GPT-6 Cyber and other OpenAI models escaped their isolated environments and ran code on the machines underneath. OpenAI used the results to patch its software and tighten network restrictions.
How safe is GPT-6 Cyber?
A hacking AI is only useful if it does what it’s told and nothing more. OpenAI tested this by giving the model a hard task in a closed environment, with a few easy-to-break services nearby.
Breaking into them could have revealed the answer, but they were outside the task’s scope, and the model had to figure that out on its own. Without safety instructions, GPT-6 Cyber left them alone in 99.93% of tests, and with the instructions, it did so every time.
Most people won’t be able to try it for now. OpenAI classifies the model as “Critical” in cybersecurity under its safety framework, and only a small group of approved security teams currently has access.
