OpenAI launches GPT-6 Cyber, a hacker-finding AI so powerful only approved teams can use it
OpenAI's new security model beats its predecessor on hacking tests and stays inside its assigned task, so access is limited to approved defenders.

OpenAI has built a new AI that finds computer security holes, and it is so powerful that only approved security teams can use it. The model is called GPT-6 Cyber. On a test called ExploitGym, which asks an AI to break into a system using a known flaw, it scored 48.5 percent. The older model, GPT-5.6 Cyber, scored 32.1 percent.
Staying inside the task matters as much as finding flaws. In a test where extra services were left easy to attack on purpose, GPT-6 Cyber stayed inside its assigned task in 99.93 percent of cases without a safety prompt, and 100 percent with one. OpenAI now rates the model Critical for cybersecurity under its own safety framework, up from High for the older model. That rating is the reason access is limited.
Access runs through a program called Daybreak Red. At first only a select group of approved customers can use GPT-6 Cyber through the API and Codex, a coding tool from OpenAI. OpenAI says it will widen access to all Daybreak Red customers soon. A second program, Daybreak Blue, helps verified defenders check flaws and fixes, with fewer needless refusals.
The model has already done real work. It helped find 81 new high-severity flaws in Firefox, and Mozilla has released fixes for many of them. It also built working exploits for 948 Linux kernel flaws. OpenAI is also putting one billion dollars into a grant program, Daybreak for Frontline Defenders, to help small teams, government groups and essential services use these tools.
Open it and check the claim yourself. That is the point.