GPT-6 Cyber found 81 serious flaws in Firefox and built 948 test attacks on Linux

When it launched GPT-6 Cyber, OpenAI mainly said that its new security model is better at finding flaws and better at staying within its task. Now it is clear what that achieved in practice with two programs that run almost everywhere: the Firefox browser and the Linux operating system. The model searched them for errors so that they can be fixed before malicious actors find them.

Firefox: 81 flaws reported, many already fixed

According to OpenAI, GPT-6 Cyber found 81 previously unknown flaws in Firefox. Mozilla, the organisation behind the browser, rated all of them 'high' for severity. For each flaw, OpenAI supplied tests with which Mozilla could reproduce the problem itself, and it proposed fixes. The model also showed that 30 already known problems were more serious than thought.

Mozilla has already released a fix for many of the flaws. In its security advisory for Firefox 159, published on 27 October, OpenAI is credited as the reporter of dozens of fixed flaws. Firefox engineer Brian Grinstead says the model helped to get fixes to millions of users. Anyone who updates Firefox receives those fixes.

Linux: 948 working test attacks

For Linux, OpenAI looked at the kernel, the part of the system that controls everything else. In a campaign in October, the model came up with 2,308 possible flaws. Of these, 1,724 were reproduced in an isolated environment.

For 948 flaws, the model built what is known as an exploit: a piece of code that proves a flaw can really be abused. These were flaws that let an ordinary user gain full control of a computer. Of those 948 flaws, 588 were previously unknown. The older model GPT-5.6 Sol reached 166 in the same set-up, so GPT-6 Cyber managed more than five times as many. The work cost OpenAI about 1.2 million dollars in computing time (about 1.1 million euros), or more than 1,200 dollars per exploit.

OpenAI reported 304 critical findings to the maintainers of Linux, with proposals for fixes. Within ten days, 61 of them had been fixed and fixes for another 43 were ready. Work on the other flaws is continuing. Analyst Sakshi Grover of research firm IDC told the trade publication CSO what matters now: how short the time is between finding a flaw and fixing it on real systems.

Stronger, but with clear limits

On ExploitGym, a test of whether a model can abuse a given flaw remotely, GPT-6 Cyber scored 48.5 per cent. Its predecessor GPT-5.6 Cyber reached 32.1 per cent. Security company Irregular, which ran a separate test with OpenAI, also saw the limits: no attack succeeded against fully secured targets, and the 15 hardest tasks remained unsolved.

That the model follows its instructions is just as important. In a test with easy-to-crack services that were not part of the task, it left them alone in 99.93 per cent of cases. Its predecessor did so in 94.49 per cent of cases. With extra safety instructions, the figure was 100 per cent. Chris Hughes of security company Noma Security told CSO this is a real improvement, while also pointing out that even a small percentage means actions on systems for which nobody gave permission.

Security company Dragos, which specialises in industrial systems, tested the model in a replica of a power plant. It found two unknown flaws there in software for managing protection relays, devices that protect a power grid during faults. OpenAI used that test to tighten its safeguards: the model now blocks risky help with industrial control, essential services and election systems even more strictly.

Under its own safety rules, OpenAI gives GPT-6 Cyber the label 'Critical', one level higher than the 'High' of its predecessor. For now, therefore, only a select group of approved security customers in the Daybreak Red programme has access. For defenders with limited resources, such as governments and providers of essential services, there is a grant programme of 1 billion dollars (about 890 million euros) that OpenAI announced earlier.

What does this mean for the road to the singularity?

GPT-6 Cyber shows that an AI model can search large, mature software, demonstrate flaws and propose fixes. That is a step towards AI that works independently on technical research, even though well-secured targets remain out of reach. The next step lies in repairing: the faster the fixes reach users, the more this power helps defenders.

What is the technological singularity? Read the explanation →