GPT-6 Cyber: OpenAI's hacking-defence AI gets its first 'Critical' rating

OpenAI launched GPT-6 Cyber on Thursday, 5 November, a version of its newest model built for security work, and gave it a “Critical” cybersecurity rating under its own Preparedness Framework. It is the first time the company has put a released model in that top risk band; its previous security model, GPT-5.6 Cyber, was rated “High”.

The rating is OpenAI saying, in effect, that the model is good enough at finding and exploiting software weaknesses that it could do serious damage in the wrong hands. So it is not on ChatGPT.

Who gets it

GPT-6 Cyber is going first to a small group of vetted customers through Daybreak Red, OpenAI’s access tier for authorised security work. The company says it will widen access to all approved Daybreak Red customers over the coming weeks. OpenAI pitches it as a tool for defenders: finding holes, testing whether they can really be exploited, and fixing them before attackers get there.

What OpenAI says it found

The company published a list of results, which have not been independently checked:

• 81 previously unknown flaws in Firefox, the web browser.
• Working exploits for 948 Linux kernel vulnerabilities, of which 588 were previously unknown.
• 304 critical Linux findings passed to the people who maintain Linux, with suggested patches.
• A score of 48.5% on the ExploitGym test, against 32.1% for GPT-5.6 Cyber, according to reports.

OpenAI also said it has seen more attempts to misuse AI, including by malicious hacking groups and by groups trying to copy (“distil”) its models.

Our take

This is the uncomfortable centre of AI security right now. The same skill that lets a model find 81 Firefox bugs for Mozilla to fix would let an attacker find them first. OpenAI’s answer is gatekeeping: vetted users, a separate tier, and handing findings to maintainers. Rivals are going the same way; Google released its own security model, Gemini 4 Carbon, last month.

For Pakistan, two things matter. Banks, telecoms and government systems here run on the same Linux and browser software, so faster patching worldwide helps them too, as long as local IT teams actually install updates. And for Pakistani security freelancers and bug-bounty hunters, the bar is moving: AI tools are starting to find the easy and medium bugs, and the human work shifts towards checking, prioritising and fixing what they find.

What to watch: whether outside researchers confirm OpenAI’s numbers, and how strictly Daybreak Red access is checked as it opens up.