OpenAI's GPT-6 Cyber Is Rated Critical, and It Found 81 Firefox Bugs Nobody Knew About

GPT-6 Cyber Is OpenAI's First Critical-Rated Security Model

OpenAI launched GPT-6 Cyber on November 5, 2026, and the headline number is hard to ignore: on ExploitGym, a benchmark that measures whether a model can turn a known vulnerability into a working exploit, GPT-6 Cyber scored 48.5%. Its predecessor, GPT-5.6 Cyber, scored 32.1%. That is a jump of more than 16 points in a single generation, on a test where every point means real exploit code that actually runs.

The model also carries a label OpenAI has never applied to one of its own systems before. Under the company's Preparedness Framework, GPT-6 Cyber is classified Critical in cybersecurity. That is the top rung of the framework, reserved for capabilities OpenAI itself judges risky enough to need special handling. The company is, in effect, shipping a tool it also flags as dangerous. That tension is the whole story.

What the Model Actually Did

The benchmark scores are only half the picture. According to reporting by The Verge, GPT-6 Cyber helped find 81 previously unknown vulnerabilities in Firefox. That is not a lab demo. Firefox is one of the most heavily audited codebases on the planet, and 81 fresh flaws in it means the model is finding things human reviewers missed.

OpenAI also says an autonomous pipeline built around the model validated and generated patches for 948 Linux local privilege escalation vulnerabilities in a single run of about 30 hours, at roughly $1.2 million in API cost. Read that sentence twice. A month ago, finding and patching 948 root bugs would have been a multi-team, multi-month project. Now it is a 30-hour job with a seven-figure compute bill. The economics of vulnerability research just changed, and every security team on earth has to reckon with it.

Two caveats belong here. First, these figures come from OpenAI and press coverage of them, not from independent verification. Second, the Linux patches are proposals for maintainers, not accepted fixes. A patch that validates in a test harness still has to survive review by the people who maintain the code. But even with those caveats, the direction of travel is clear.

Why the Critical Rating Matters

OpenAI's Preparedness Framework exists to decide when a capability is too risky to ship normally. A Critical rating means the company believes the model could materially help with cyberattacks if misused. The fact that it shipped anyway, with safeguards, tells you how OpenAI is thinking about this: the defensive value of the model outweighs the offensive risk, at least for now.

There is a real argument on both sides. Defenders get a tool that finds bugs faster than any human team. Attackers get the same tool. The difference is who moves first and who patches fastest. That is why the Firefox result matters more than the benchmark score. Finding 81 unknown bugs and getting them fixed before anyone else exploits them is the model doing exactly what its defenders hoped. The risk is that the same capability, in the wrong hands, finds the next 81 first.

This is the same tension the industry has lived with since the first automated fuzzers. GPT-6 Cyber just raises the stakes by an order of magnitude. The question is no longer whether AI can find vulnerabilities. It can. The question is whether the patching pipeline can keep up with the discovery pipeline, and right now that is an open question.

What This Means for Security Teams

If you run a security team, the practical takeaway is blunt: the window between a vulnerability existing and someone finding it is shrinking fast. The old rhythm of quarterly audits and annual penetration tests assumed discovery was slow and expensive. GPT-6 Cyber makes discovery cheap and fast, which means your patching cadence is now the bottleneck.

Three things are worth doing now. First, shorten the time between a patch being available and a patch being deployed, because the discovery side just got a lot faster. Second, assume that any bug a model can find in your code, an attacker with the same model can also find, and prioritize accordingly. Third, start using these tools yourself. The Inkling 2 Small launch showed how quickly capable models are reaching developers, and the security field is no exception. The teams that adopt AI-assisted auditing early will find their own bugs before someone else does.

GPT-6 Cyber is not the end of this story. OpenAI teased GPT-5.6 Sol after briefing US officials, and the pace of releases suggests a GPT-6 family that keeps getting more capable at specific jobs. The GPT-6.2 Astra leaks hint at what comes next, though much of that remains unverified. What is verified is that the model that shipped this week is already good enough to find real bugs in real software, at a scale humans cannot match. Security work will never go back to the old pace.