OpenAI has given its newest security model the highest cybersecurity risk rating it has ever assigned to one of its own systems. GPT-6 Cyber, released on Thursday, November 5, was rated "Critical" under the company's Preparedness Framework after it found 81 previously unknown high-severity flaws in Mozilla Firefox and built working exploits for 948 vulnerabilities in the Linux kernel, according to OpenAI's own test results.
 |
| File photo: Signage at an OpenAI office. OpenAI released GPT-6 Cyber on November 5, 2026, and rated it 'Critical' for cybersecurity under its Preparedness Framework. |
The release marks a striking moment for the artificial intelligence industry: a leading AI lab is openly saying one of its models is powerful enough at hacking to require locked-down access, even as it argues the same capability can help defenders find and fix flaws faster than humans ever could. OpenAI is not making the model widely available. For now, only a small group of vetted security researchers and enterprise teams can use it.
What GPT-6 Cyber Is
GPT-6 Cyber is a specialised version of OpenAI's GPT-6 Astra model, tuned for authorised cybersecurity work. According to OpenAI, it is built to discover previously unknown vulnerabilities, reproduce system crashes, write working exploit proofs-of-concept, and check whether patches actually fix the flaws they target.
The model is being offered through OpenAI's API and its Codex coding tool to a select group of approved customers in a trusted-access programme called Daybreak Red, with access expanding to all Daybreak Red customers over the coming weeks. OpenAI has not said when — or whether — the model will be made more widely available.
The "Critical" rating is the top tier of OpenAI's Preparedness Framework, the internal system the company uses to judge how dangerous a model could be in the wrong hands. The previous security model, GPT-5.6 Cyber, was rated "High". GPT-6 Cyber shares its other Preparedness ratings with GPT-6 Astra: "High" for biological and chemical risk, and below "High" for AI self-improvement.
The Numbers Behind the 'Critical' Rating
The headline figures behind the rating come from OpenAI's own testing, and the company has been clear that they have not been independently verified. On ExploitGym, a benchmark that tests whether a model can exploit a specified vulnerability, GPT-6 Cyber scored 48.5 percent, compared with 32.1 percent for GPT-5.6 Cyber and 42.4 percent for GPT-6 Astra.
The Linux kernel results are the most dramatic. In a mid-October autonomous run, OpenAI says the model reviewed 2,308 candidate findings, reproduced 1,724 vulnerabilities, and built working local privilege-escalation exploits for 948 distinct flaws — 588 of them previously unknown — at a cost of about 1.2 million dollars in API credits, or roughly 1,279 dollars per exploit, in about 30 hours. The same research setup produced just 166 working exploits from the earlier GPT-5.6 Sol model. OpenAI says it disclosed 304 critical findings to Linux kernel maintainers, each with a candidate patch.
The Firefox results carry an important difference: they were confirmed by an outside party. OpenAI says the model produced evidence and reproducible tests for 81 previously unknown vulnerabilities that Mozilla classified as high severity, and showed 30 known issues were worse than reported. Mozilla documented the flaws in its own security advisory, MFSA 2026-118, and fixes are now landing. OpenAI also says the model completed 95 percent of exploit-development tasks on its new RemoteExploitBench test.
OpenAI also reports the model is far better at staying within the limits researchers set. In honeypot tests, it went out of scope in just 0.07 percent of runs without deployment safety instructions, compared with 5.51 percent for its predecessor.
Why OpenAI Is Keeping It Behind Closed Doors
The gating is the point. A model that can build a working exploit for a previously unknown flaw for about 1,279 dollars is, by definition, a dual-use tool: the same capability that helps a security team find a bug before criminals do could help an attacker find it first. OpenAI's answer is to limit access to vetted defenders — security researchers and enterprise red and blue teams — and to keep the model out of general release.
The company has also set pricing that signals the model is aimed at professional security work rather than casual use: 20 dollars per million input tokens, 2 dollars per million cached input tokens, and 100 dollars per million output tokens, according to reports citing VentureBeat.
The release lands in a week when the AI industry's safety promises are under unusual scrutiny. Anthropic, OpenAI's main rival, is preparing a stock market debut that investors have pushed to a lower valuation after a sharp sell-off in technology shares, and the company warned in its own filings of what it called catastrophic AI risks. OpenAI's decision to publish detailed capability numbers — while keeping the model itself restricted — is part of a wider argument the labs are making: that the risks of powerful AI are real, measurable, and manageable only with controls.
 |
| File photo: A Linux terminal showing system directories. OpenAI says GPT-6 Cyber built working exploits for 948 Linux kernel vulnerabilities, including 588 flaws that were previously unknown. |
What This Means
The most important number in the GPT-6 Cyber announcement may not be 948. It is 81 — the Firefox flaws that Mozilla itself confirmed. The Linux kernel figures are impressive, but they come only from OpenAI, with no public list of vulnerabilities for outside researchers to check one by one. The Firefox results show the capability is real in at least one major, independently verified codebase, and they show the model working the way defenders hope: flaws found, disclosed, and patched.
The cost figures matter just as much. If a model can produce a working exploit for roughly 1,279 dollars in computing time, the economics of vulnerability research change. Work that once took teams of specialists weeks can, in principle, be done in hours. That is good news for the defenders who find flaws first — and a warning about what happens if the same tools reach attackers who do not disclose what they find.
The gap between the two predecessors is also striking. GPT-6 Cyber produced more than five times as many working Linux exploits as GPT-5.6 Sol under the same research conditions, and its ExploitGym score jumped by more than 16 percentage points. That pace of improvement is why OpenAI raised the risk rating: the company's own framework says a model that crosses this threshold needs stronger safeguards, and the rating is the mechanism that triggers them.
There are real limits to what the announcement proves. Fewer than one in ten of the Linux flaws had fixes merged upstream ten days after the run, according to independent analysis of the launch, which suggests that finding vulnerabilities is easier than getting them fixed. And the model was tested in sandboxes against reproduced flaws — research results, not attacks in the wild.
 |
| File image: The Firefox logo. Mozilla confirmed 81 previously unknown high-severity flaws found by the model and documented them in security advisory MFSA 2026-118. |
What Happens Next
The immediate test is whether the model's findings keep turning into real fixes. Mozilla is already shipping patches for the 81 Firefox flaws, and Linux maintainers are working through the 304 critical findings OpenAI disclosed with proposed patches. How quickly those fixes land — and whether other software projects invite the same kind of AI-assisted auditing — will show whether the model is a practical defensive tool or mainly a demonstration.
The next question is access. OpenAI says Daybreak Red access will widen over the coming weeks, but it has given no date for broader availability. Security researchers will be watching whether the gating holds, and whether other AI labs follow OpenAI in publishing capability numbers for their own security models.
The longer question is the one OpenAI's own rating raises. If each new generation of AI is several times better at finding and exploiting software flaws, the window between a vulnerability being discovered and being fixed becomes the critical battleground. GPT-6 Cyber is rated "Critical" — the top tier of OpenAI's risk scale — and the company is betting that keeping it in defenders' hands is the way to make software safer, not riskier.
Post a Comment