GPT-6 Cyber: what OpenAI's 'Critical' security model does
Key takeaways
- GPT-6 Cyber is restricted to vetted Daybreak Red customers, not sold to the public
- OpenAI says it found 81 high-severity Firefox flaws and built exploits for 948 Linux kernel vulnerabilities
- The rating rose from 'High' to 'Critical', and OpenAI reports more attempts to misuse AI
- Defenders should expect faster vulnerability discovery and shorter patch windows
What does it mean when an AI company rates its own model "Critical" for cybersecurity, and still ships it? OpenAI has done exactly that with GPT-6 Cyber, a model built to find and exploit security flaws, released on Thursday to a small group of vetted customers. The rating was raised from 'High', and the access controls around the model are as much a part of the story as the model itself.
What GPT-6 Cyber actually is
GPT-6 Cyber is a version of OpenAI's model aimed at one job: helping security researchers find vulnerabilities, test how they could be exploited and fix them before attackers get there. An exploit is the working code that abuses it.
The model is not on general sale. It is offered through Daybreak Red, which OpenAI describes as its access tier for advanced, authorised security work such as penetration testing and exploit development. Penetration testing means hiring people to attack your own systems under agreed rules so the weaknesses surface in a controlled setting.
OpenAI has raised the model's cybersecurity classification to "Critical" under its safety framework, from "High" for GPT-5.6 Cyber. The company says it is seeing more attempts to misuse AI, including by malicious cyber actors and groups seeking to distill its models. Distillation means training a cheaper model to imitate a more capable one by studying its outputs.
How it works in practice
The model reads code, spots where logic breaks down, and then goes a step further by writing the paperclip: a working exploit that proves the flaw is real. That second step matters because a proven exploit tells maintainers a bug is urgent rather than theoretical.
OpenAI's own figures give a sense of the results. Its researchers used the model to find 81 previously unknown Firefox vulnerabilities that Mozilla classified as high severity, and Mozilla has begun releasing fixes. On the Linux kernel, the core of the operating system behind most servers, the model helped develop working exploits for 948 vulnerabilities, 588 of which were previously unknown. OpenAI has also disclosed 304 critical Linux findings to maintainers, along with proposed patches.
Those numbers are OpenAI's claims about its own research; the Firefox figure carries independent weight because Mozilla did the severity classification.
Who uses it and why
The intended users are the people who already do this work by hand: penetration testers, internal red teams (staff paid to attack their employer's systems) and vulnerability researchers. For them, a model that triages code and drafts proofs of concept compresses days of effort. The disclosure of proposed patches alongside the Linux findings points to the other half of the loop, where the same tool helps maintainers close what it opens.
Open-source projects are the obvious beneficiaries and the obvious pressure point. Mozilla and the Linux maintainers receive findings from OpenAI that they must triage and fix, and the 304 critical Linux disclosures add to what maintainers must work through.
What the limitations are
The first is the dual-use problem. A model that can write 948 working kernel exploits is, by definition, useful to attackers, which is why OpenAI gates it behind vetting and why it carries the top rating. Vetting reduces the risk; it does not remove it.
The second is access. OpenAI says GPT-6 Cyber reaches select Daybreak Red customers first and will expand to all approved customers over the coming weeks. Smaller security firms and in-house teams outside that programme get no direct use of it for now. OpenAI is making GPT-5.6 Cyber available to all Daybreak Red customers from Thursday, so there is a lower rung for everyone approved.
The third is that finding flaws is faster than fixing them. Discovery scales with compute; patching still depends on human maintainers, testing and release cycles.
What this means in practice
For UK and US security teams, the practical effect is on timing. If vetted defenders can surface and prove flaws at this rate, the gap between a bug existing and a bug being known shrinks, which may put pressure on organisations to patch sooner. Teams that track actively exploited flaws, for example through the CISA catalogue, will want to check the CISA KEV catalogue as part of routine patch triage rather than as an occasional exercise.
Neither document responds to this launch specifically; the source reports no UK regulatory action.
Procurement is the other practical point. A team considering a Daybreak Red subscription should expect vetting and should ask how findings, including exploit code, are stored and shared.
Where this is heading
OpenAI's own wording points the direction: a higher rating than the previous generation, more misuse attempts, and access widening in stages. The pattern is likely to repeat, with each model generation finding more flaws, restricted tiers gating the most capable versions, and defenders told to patch faster. Watch three things: how quickly Mozilla and the Linux maintainers work through the disclosed findings, whether access beyond Daybreak Red customers is ever offered, and whether other AI developers publish comparable figures for their own security models.
Frequently asked questions
What is GPT-6 Cyber?
A model from OpenAI built to find security flaws, test how they could be exploited and help fix them. It went to a small group of vetted customers on Thursday.
Who can use GPT-6 Cyber?
Select customers of Daybreak Red, OpenAI's access tier for authorised security work such as penetration testing and exploit development. Access widens to all approved Daybreak Red customers over the coming weeks.
What does the 'Critical' rating mean?
OpenAI raised GPT-6 Cyber's cybersecurity classification to 'Critical' under its safety framework, from 'High' for GPT-5.6 Cyber.
Is GPT-5.6 Cyber also available?
Yes. OpenAI is making the earlier GPT-5.6 Cyber model available to all Daybreak Red customers starting Thursday.