AI Cybersecurity: OpenAI Rated Its Own Model Critical

AI Cybersecurity: OpenAI Rated Its Own Model Critical

OpenAI’s GPT-6 Cyber Is Rated Critical for AI Cybersecurity

The model finds flaws and builds exploits at a scale that has no precedent. The question the launch cannot answer is whether vetted access is a control or a formality.

OpenAI has released GPT-6 Cyber, a security-focused model it rates Critical for cybersecurity under its Preparedness Framework, the highest risk category the company defines and one step above the High rating given to GPT-5.6 Cyber. Access runs through Daybreak Red, OpenAI’s vetted tier for authorised security work, and the model is not publicly available: organisations need separate approval, stronger verification and additional access controls before deploying it through Codex or the Responses API. The capability claims are specific enough to examine. What the release cannot answer is whether that vetting amounts to a real control.

What Critical actually means

The Preparedness Framework sorts advanced systems by the risk they pose across cybersecurity, biology, chemistry and AI self-improvement. A Critical cybersecurity rating, as the framework defines it, covers a model that can identify vulnerabilities, develop functional zero-day exploits against real critical systems and execute end-to-end cyberattacks.

That describes capability rather than intent, and the distinction carries weight. OpenAI is not saying the model is hostile; it is saying the model can do work that previously required a skilled attacker. A rating that reads as alarming is, in framework terms, a classification triggering deployment restrictions rather than a warning about the product’s character.

The comparison supplies the significance. GPT-5.6 Cyber sat at High. Moving to Critical means OpenAI’s own evaluation concluded the capability crossed a threshold, and the company publishes that conclusion about its own product while shipping it to a restricted group.

Firefox: 81 flaws and the patch pipeline

The Mozilla result is the most concrete evidence in the release. OpenAI says the model identified 81 previously unknown high-severity vulnerabilities in Firefox, a browser earlier frontier models had already tested extensively.

Three details lift that above a headline number. The model produced evidence and reproducible test cases for each finding rather than flagging suspicious code. It proposed patches. And Mozilla classified all 81 as high severity and began releasing fixes shortly after receiving the reports, which is the closest thing to third-party corroboration in the document, since a maintainer’s severity classification is an independent judgement.

The model also found that 30 already-known Firefox vulnerabilities were more severe than previously assessed. That is the quieter and arguably more useful result. Severity misclassification causes organisations to defer remediation, and a tool correcting it improves an existing process rather than only adding to the queue.

Linux: 948 exploits and the disclosure question

The kernel results carry more weight and more risk. OpenAI says GPT-6 Cyber reproduced 1,724 vulnerabilities and developed working exploits for 948 distinct Linux kernel flaws, primarily local privilege escalation, meaning an attacker with limited access could gain root.

The scale comparison is the striking part. Using the same research harness, GPT-5.6 Sol produced 166 working exploits. GPT-6 Cyber produced 948, more than five times as many.

OpenAI says it reported 304 critical findings to Linux maintainers with candidate patches and is working to get the remainder fixed.

That sentence deserves more scrutiny than the release gives it. Disclosing 304 critical findings to a volunteer-maintained project is a substantial load, and the release does not say over what period, whether maintainers asked for the volume, or how many have been fixed. Kernel security work runs on a small number of people with a structured disclosure process, and a model generating findings faster than humans can triage them creates a queue problem rather than a security improvement. The release treats reporting as the end of the responsibility; it sits closer to the beginning.

There is also a category question in the Linux numbers that the release leaves open. Reproducing 1,724 vulnerabilities and building working exploits for 948 of them describes a capability demonstration, not a disclosure programme. The two are different outputs. A demonstration tells the sector what AI cybersecurity models can now do. A disclosure programme tells maintainers what needs fixing. OpenAI reports the first in detail and the second only as a count of 304, so a reader cannot tell how much of the 948 is already being addressed.

The disclosure norm

The release does not discuss the disclosure norm it operates inside. Coordinated disclosure traditionally gives a vendor a defined window to fix a flaw before publication, and the ethical obligation sits with the researcher who found it. When a model finds 81 flaws in one browser and 948 exploitable defects in a kernel, the volume changes the shape of that obligation. A maintainer facing a batch that size cannot triage it the way a single report is triaged, and the researcher’s duty to disclose responsibly becomes a duty to pace the disclosure so the receiver can absorb it. Nothing in the release suggests OpenAI has addressed pacing, and for AI cybersecurity vendors this is the operational question that will define the next two years.

The 99 per cent problem

OpenAI describes a scope test more revealing than its framing suggests. Researchers gave the model an exploitation task alongside supporting services that were deliberately easy to compromise but explicitly out of scope. The model avoided exploiting those services in more than 99 per cent of cases.

The release calls that encouraging and then concedes the obvious: more than 99 per cent is not 100, and OpenAI has not detailed the cases where the model crossed its assigned boundary.

For a system rated Critical, the tail is the story. A model staying in scope in 99 of 100 runs and developing a working exploit outside it in the hundredth is a model whose failures matter more than its successes, because the failure mode is unauthorised access rather than a bad answer. The release does not say how many runs the test involved, which determines whether the tail is one case in a hundred or one in a thousand. That single number separates a rounding error from a deployment blocker in practice.

The market and the vetted-access model

OpenAI is not alone here. The release notes that Anthropic and Google compete for security-focused AI customers through their own vetted-access programs, and cites a Gartner forecast from September that worldwide spending on AI cybersecurity would reach 51.3 billion dollars in 2026.

That figure explains the commercial logic. Vetted access is not only a safety measure; it is a market structure. Restricting a model to approved researchers creates a tier governments and enterprises will pay for, and it gives the vendor a defensible position against open-weight competitors who cannot offer the same assurance.

Whether the structure works depends on what vetting means in practice. The release describes separate approval, stronger verification and additional access controls without saying who approves, against what standard, or what happens to an organisation that loses approval. A tier whose entry criteria stay unpublished is a commercial filter with a safety narrative attached. It may be both, and the release does not draw that distinction anywhere.

What authorised work actually means

Authorised security work carries a lot of load in this announcement and the release does not define it. The categories it names, red-teaming, penetration testing, exploit-chain validation and controlled vulnerability research, are legitimate professional activities carried out by organisations holding contractual authority over the systems they test. That authority separates a penetration tester from an attacker, and it is a legal and contractual boundary rather than a technical one. A model cannot verify it. The access controls OpenAI describes, identity requirements, check-ins and monitoring, exist to establish that the organisation behind the request holds that authority. Whether they succeed depends on the strength of the verification, which the release does not describe.

The benchmark problem

One more gap deserves attention: the benchmark itself. ExploitGym measures whether a model can exploit a specified vulnerability in a remote target, and OpenAI reports 48.5 per cent against 32.1 per cent for its predecessor, a relative improvement of about 51 per cent. That is a large jump on a task designed to be hard, and no third party has reproduced it. For AI cybersecurity buyers, a benchmark score is the only portable comparison between vendors, and a score only the vendor reports stays a marketing number until someone else runs the harness. The same applies to the exploit counts: 948 becomes a sector reference point only after independent confirmation.

What is missing

No publication date appears anywhere in the document, which for a launch is a disclosure failure. No statement from Mozilla or from Linux kernel maintainers confirms the findings, the severity classifications or the patch status. The release does not say how many organisations hold Daybreak Red access, what the approval standard is, or how the small share of out-of-scope failures breaks down. No independent evaluation of the exploit claims appears, and no third party has reproduced the ExploitGym result.

AI Cybersecurity: OpenAI Rated Its Own Model Critical
Photo by Levart_Photographer on Unsplash

What to watch

Whether Mozilla and kernel maintainers confirm the findings in their own words, which would convert vendor claims into verifiable outcomes. Whether OpenAI publishes the out-of-scope failure rate alongside a sample size. And, whether the 304 Linux findings translate into fixes at a rate maintainers consider sustainable, because a disclosure pipeline outrunning triage is a new kind of problem. Whether Daybreak Red’s approval criteria become public. And whether the Critical rating holds as access widens, since a restriction that expands to all approved customers within weeks is a restriction only if the approval process has teeth behind it.