OpenAI has released an experimental model, GPT-6 Cyber, and classified it as "Critical" for cybersecurity under its own internal safety framework. Its predecessor, GPT-5.6 Cyber, was rated "High".
The rating means the model's cybersecurity capabilities, and its potential for misuse, call for extra caution. It is the company grading its own release, and grading it upward.
What it is built to do
The model is aimed at cybersecurity researchers. It finds vulnerabilities in software and systems and tests how they could be exploited, so that defenders can fix them before criminals reach them.
That is the oldest bargain in security research, and it has never been comfortable. The same work that lets a defender patch a flaw tells an attacker where the flaw is. What changes with a tool like this is not the bargain but the speed.
The results are the argument, both ways
OpenAI researchers used it on the Mozilla Firefox browser and found security flaws that Mozilla rated high-risk; updated versions have begun to appear.
More striking: 304 serious security flaws were reported to the maintainers of the Linux kernel, with proposed patches attached.
Three hundred and four is a number worth sitting with. The Linux kernel is among the most heavily reviewed pieces of software ever written, maintained by a large and experienced community, and running on most of the world's servers and nearly all of its phones. If a model can surface that many serious issues in it, the honest reading is not that the kernel is badly made. It is that human review at scale has limits this tool does not share.
There is a second problem inside that figure, and the maintainers will feel it before anyone else does. Three hundred and four reports with patches attached is also three hundred and four patches that a person has to read, understand and judge before anything is merged. Kernel review is a bottleneck made of people, and a tool that produces work faster than people can absorb it does not automatically make the software safer — it can just as easily move the queue from finding flaws to approving fixes.
Who can use it
Access is initially restricted to selected customers in OpenAI's "Daybreak Red" programme, described as being for advanced, authorised cybersecurity practitioners.
That is the containment, and its weakness is obvious: it is a commercial allowlist, not a technical control. It works for as long as the capability stays inside it. Every previous capability of this kind has eventually been matched by something with no allowlist at all.
What it means for anyone running a server
The practical conclusion is unglamorous and urgent. If flaws in widely used software can now be found this quickly, the window between a vulnerability becoming known and being exploited gets shorter — and the only defence that scales is applying patches promptly.
For the many Bangladeshi businesses running their own Linux servers, often with updates deferred because a reboot is inconvenient, that is the whole lesson. The kernel patches coming out of this work protect you only once they are installed.
