OpenAI just released an AI model that is, by its own internal measure, too dangerous to hand out freely. GPT-6 Cyber found 81 previously unknown high-severity flaws in Firefox and built 948 working exploits for Linux kernel bugs during testing, and OpenAI rated the model "Critical for cybersecurity," the highest tier in its risk framework. That is not marketing language. It is the same classification system OpenAI uses to decide who is allowed anywhere near the model.
What the model actually did
GPT-6 Cyber is a specialized version of OpenAI's Astra model line, trained specifically for authorized vulnerability research rather than general chat or coding tasks. In testing, it was pointed at 5,034 audit targets in the Linux kernel and came back with 948 exploits that actually worked, plus it flagged 30 known Firefox issues as more severe than originally reported. Mozilla confirmed and disclosed the Firefox findings in a security advisory.
On ExploitGym, a benchmark that measures whether a model can exploit a specified vulnerability in a remote target, GPT-6 Cyber scored 48.5%, up from 32.1% for the previous GPT-5.6 Cyber model. That is a large jump in a single generation, and it is the kind of number that security teams either find reassuring or alarming depending on which side of the fence they sit on.
The part that is easy to miss: cost
The Linux exploit run reportedly cost OpenAI around $1.2 million to produce, which works out to roughly $1,279 per working exploit. Using historical bug bounty pricing, OpenAI estimated the value of those 948 exploits at somewhere between $18.9 million and $66.3 million.
That price comparison matters more than it looks. Finding and weaponizing a kernel exploit used to require scarce, highly paid human talent. If a model can do it for roughly a thousand dollars a pop, the economics of who can afford to go looking for exploits just changed, for defenders and attackers alike.
Why OpenAI called it Critical, not just powerful
OpenAI's Preparedness Framework rates models across risk categories, and GPT-5.6 Cyber had been rated "High." GPT-6 Cyber cleared the bar into "Critical," the top tier, specifically because of its offensive cyber capability. That is a meaningful admission: the company is saying, in effect, that its own model is now capable enough to need a different set of guardrails than the version that came before it.
The tension here is real and OpenAI does not fully resolve it. A model that is good enough to find serious bugs before criminals do is genuinely useful to security teams. The exact same model, in the wrong hands or used outside its intended scope, is a tool for building attacks faster than defenders can patch them. OpenAI's own safety writeup frames it bluntly: the risk with a model like this isn't only whether it works, it's whether it wanders outside the boundaries it was authorized for.
How access actually works
GPT-6 Cyber is not available to the general public or even to all paying OpenAI customers. It ships only through a program called Daybreak Red, which requires separate approval on top of OpenAI's standard API access, plus mandatory single sign-on and multi-factor authentication on the customer's side.
- Access is limited to vetted security teams doing authorized vulnerability research, exploit validation, or red-teaming work.
- Requests into the API must come from pre-approved networks by default.
- OpenAI layers in cyber-specific misalignment monitoring and restrictions around anything touching critical infrastructure.
- Sensitive workflows are expected to run with human oversight, not fully autonomously.
On the safety numbers OpenAI did publish, the improvement is notable. Without any deployment safety instructions, earlier cyber models stayed within their authorized scope about 94.5% of the time. GPT-6 Cyber reportedly improved that to around 99.9% even with no safety prompt at all, and effectively 100% when given one. In other words, OpenAI says the model got both more capable and better at staying inside the lines it was given, at the same time.
Why this matters beyond one model
This is part of a broader pattern worth watching, not a one-off release. AI companies are increasingly building specialized "offensive capability" models for security work, and every one of them faces the same fork: the more useful the model is for finding bugs before attackers do, the more dangerous it is if that capability leaks or gets misused. There is no clean technical fix for that, only tighter access controls and hoping the controls hold.
It is also a preview of where AI-assisted security work is heading generally. If a model can turn up nearly a thousand working kernel exploits in one run, the bottleneck in both offense and defense is shifting away from "can someone find the bug" and toward "who gets to run the model that finds it, and what do they do next."
The takeaway
The interesting story here is not that AI can find security bugs, researchers have used machine learning for vulnerability hunting for years. It is that OpenAI built something capable enough that the company felt compelled to rate its own product at the top of its danger scale and then lock it behind an approval process most of its customers will never pass. That is a rare moment of a company treating its own release as a genuine risk decision rather than a straightforward product launch, and it is worth watching whether that caution holds as the next version gets even better at the same job.