GPT-6 Cyber found 81 new Firefox vulnerabilities and wrote 948 Linux exploits

Vulnerability research used to be painstaking work: an experienced researcher would spend weeks poring over code and hunting for suspicious areas, then spend just as long writing an exploit to prove that the flaw was real. OpenAI says its new GPT-6 Cyber model can perform both parts of that job. If the figures are confirmed by independent reviews, the economics of security will change significantly.

What OpenAI actually demonstrated

GPT-6 Cyber is a version of the model tuned for vulnerability discovery and the defense of computer systems. The company highlights two main results.

First, the model helped discover 81 previously unknown vulnerabilities in the Firefox browser. The word “unknown” is crucial here. These are not old CVEs recalled from the training data, but new findings in a large, long-scrutinized project. Firefox has been fuzz-tested and examined by researchers for years, so finding something new there is no easy task.

Second, the model developed 948 working exploits for the Linux kernel. That is another level entirely. A static analyzer can identify a suspicious section of code. Turning it into a working exploit that actually reproduces the problem is much harder. A working exploit settles the central debate within any team: “Can this actually be exploited?”

Why the 0.07% figure matters

OpenAI also says the model is better at complying with established restrictions. In tests, the violation rate fell from 5.51% to 0.07%.

For a cyber model, this is not a cosmetic improvement. Such a tool is inherently dual-use: the same capability that helps patch a flaw can help exploit it. The question, therefore, is not only how intelligent the model is, but how reliably it stays within the rules: refusing to help attack third-party systems and remaining within the boundaries of the authorized task.

The decline from 5.51% to 0.07% means that a violation went from occurring once in every 18 cases to roughly once in every 1,400. But 0.07% is not zero. At scale, even that rate produces hundreds of incidents across hundreds of thousands of requests. Moreover, this figure comes from OpenAI's own tests, not from the real world, where users actively search for workarounds.

What this means in practice

For developers and security teams, the takeaway is that AI-assisted code auditing is becoming a practical tool rather than a demo. If you maintain an open-source library, browser component, or low-level code, you should assume that similar models will scan your repository—and that you will not be the only one doing so.

What teams should do now:

• Speed up the remediation cycle. If findings arrive faster, a slow patch-release process becomes the main bottleneck.
• Automate reproduction. A working exploit is useful not only to attackers; it helps teams quickly confirm and prioritize a bug.
• Reassess the threat model. Flaws in old code that has long been considered “reviewed” may not be as safe as assumed.
• Develop a disclosure policy. A stream of AI-generated findings will require clear reporting channels and remediation deadlines.

Security-product teams and founders should pay even closer attention. The ability to find a bug is becoming cheaper on its own. Value is shifting to everything around it: triage, CI integration, automated fixes, and evidence that a patch actually resolves the problem.

Limitations and unanswered questions

For now, we are relying on OpenAI's claims. It is unclear how many of the 81 discovered vulnerabilities are critical and how many are low-risk. It is also unclear what the 948 Linux exploits looked like: how reliable they were, whether they depended on specific configurations, and how much human involvement they required. There is no information about the false-positive rate either, even though that determines whether the tool saves time or buries the team in noise.

The conclusion is straightforward: GPT-6 Cyber is a strong signal that AI has moved beyond code suggestions and into autonomous security-research work. Defenders should adopt it before attackers do.