GPT-6 Cyber Finds 81 Dangerous Vulnerabilities in Firefox and Scores 48.5% on ExploitGym

Until now, language models in security mostly helped read logs and explain other people's CVEs. Judging by the results OpenAI has published, GPT-6 Cyber takes the next step: it doesn't just analyze known bugs, it looks for new ones and writes exploits for them. For anyone responsible for a product or infrastructure, this is a signal: the pace of vulnerability discovery is changing.

What the model actually showed

Here are the numbers from the announcement. On the ExploitGym benchmark, GPT-6 Cyber scored 48.5%, which OpenAI calls a new record. In Firefox, the model discovered 81 previously unknown high-severity vulnerabilities. And in Linux, it showed how 948 bugs can be used for privilege escalation.

It is important to tell these three results apart. The benchmark measures the ability to turn a discovered bug into a working exploit on standard tasks. Finding new bugs in Firefox involves real code that thousands of people install on their machines. And the 948 bugs in Linux are not 948 ready-made hacks, but bugs that, according to OpenAI, the model managed to link to privilege escalation.

What it looks like in a simple example

Imagine you have an old C module that parses a user-supplied file. Previously, an auditor would spend weeks reading the code, setting up a fuzzer and looking for where a buffer goes out of bounds. Now an agent in Codex receives the repository, launches the build itself, tries inputs, sees a crash, and then works out whether that crash can be turned into execution of someone else's code or root access. It is exactly this second step, from "the bug exists" to "the bug works," that ExploitGym tests.

For defenders this is good news: what used to take weeks can be done in hours and run across the entire codebase, not just the most critical parts. For attackers the news is the same, with the opposite sign. It's no coincidence that commenters on the announcement joke that hackers will get access too.

Who gets access and what remains unclear

GPT-6 Cyber is currently available in Codex and via API only to select participants in the Daybreak Red program. This looks like a cautious release: vetted researchers first, then possibly a wider audience. The expected GPT-6.2 Astra and Sol did not launch today, so it is too early to talk about them.

The limitations are also worth keeping in mind. All the figures so far come from OpenAI's own statements, and the announcement includes no independent verification. A 48.5% benchmark score means the model still fails more than half of the tasks. It is unknown how many of the 81 findings are actually exploitable and how many false positives humans had to filter out. The cost of such a run has not been disclosed either.

What to do in practice

First: if you maintain C or C++ code, the kernel, browser components or network services, assume that AI auditing will become a routine part of the process, and not only on your side. The window between a bug appearing and being exploited will shrink.

Second: speed up your patch cycle. If updating dependencies takes you a month, that is already a risk. Automate builds, tests and the rollout of fixes.

Third: if you work in security professionally, look at the Daybreak Red program and its access terms. Early familiarity with such tools will give you an advantage before they become mainstream.

Fourth: don't trust the numbers blindly. Wait for independent verification and write-ups from the Firefox and Linux maintainers themselves: they will say how real the findings were.

The main takeaway is simple: AI in security is turning from an assistant into an independent researcher. The question is no longer whether a model will find a hole in your code, but who will do it first: you or someone else.