OpenAI releases GPT-6 Cyber, a bug-hunting AI it rates ‘Critical’ for hacking skill, to vetted security teams only
The model behind this week’s 948 Linux exploits also found 81 serious Firefox flaws, OpenAI says. Cisco is among the first users.
OpenAI on Thursday released GPT-6 Cyber, an artificial intelligence model built to find and test weaknesses in software, the company said. Only a small group of approved security teams can use it for now.
Software flaws, called vulnerabilities, are coding mistakes that attackers can use to break into computers. Researchers hunt for them so they can be fixed first. To prove a flaw is real, they often write an “exploit,” a small program that actually uses it. OpenAI says GPT-6 Cyber can do much of that work itself: read code, spot a weakness, build a working exploit and draft a fix.
On ExploitGym, a test of whether an AI can exploit a known flaw, the model scored 48.5%, up from 32.1% for the earlier GPT-5.6 Cyber, OpenAI said. The company said it found 81 previously unknown high-severity flaws in the Firefox web browser, and Mozilla, which makes Firefox, has released fixes for many of them. It was also behind OpenAI’s Linux research released Wednesday, which produced working exploits for 948 flaws.
OpenAI rates the model “Critical” for cybersecurity under its Preparedness Framework, the company’s system for grading how dangerous a model’s abilities are. The prior version was rated “High.” Because the same skills could help criminals, access is limited to Daybreak Red, OpenAI’s vetted program for professional security testers. OpenAI says it monitors use and blocks help with attacks on power plants, water systems and other critical services. In OpenAI’s tests, the model stayed within its assigned targets 99.93% of the time without safety instructions and 100% with them.
Regular ChatGPT users cannot use it, but readers benefit when flaws in common software get patched, one more reason to install updates. San Jose’s Cisco, Palo Alto Networks and Mountain View’s SentinelOne are among companies OpenAI quoted as using the model.