OpenAI’s GPT-6 Cyber detects vulnerabilities and suggests fixes
OpenAI presents GPT-6 Cyber, a model designed to detect vulnerabilities, build exploits, and suggest fixes. The company says it has already tested it on Firefox and the Linux kernel, with ambitious but independently unverified results.
OpenAI unveiled GPT-6 Cyber, an AI model specialized in finding security flaws, producing functional exploits, and proposing patches. Presented on November 5, 2026, it mainly targets security teams and authorized researchers, with limited access for now. The announcement matters for the cybersecurity market as it shows how far models can go, including on very sensitive targets like Firefox and the Linux kernel.
A model designed to attack, then defend
OpenAI presents GPT-6 Cyber as a system trained to find vulnerabilities in software, then to derive concrete exploits and patches. The company also claims to have strengthened its safeguards, with permission controls, targeted filtering of recently disclosed flaws, automatic checks via Private Safety Processing, and IP address access restrictions for Daybreak Red.
In internal tests cited by OpenAI, the model avoids perimeter breaches in 99.93% of cases without specific security instructions, compared to 94.49% for its predecessor. With security instructions enabled, the rate rises to 100%. On paper, the model also scores 48.5% on ExploitGym, a benchmark measuring the ability to exploit a remote vulnerability, versus 32.1% for GPT-5.6 Cyber. On a test for finding flaws in web applications without source code access, OpenAI reports a 40.3% success rate.
Firefox, Linux and the thorny question: help or weapon?
The most sensitive point of the announcement lies in its results on real systems. OpenAI says GPT-6 Cyber found 81 previously unknown critical vulnerabilities in Firefox, already fixed by Mozilla. The model also reportedly showed that 30 known issues were more severe than estimated. On the Linux kernel, the company claims to have reproduced 1,724 vulnerabilities and obtained functional exploits for 948 of them. These figures should be taken with caution, as they all come from OpenAI and no independent evaluation has yet been published.
The cost and scale of the setup highlight the issue. For the Linux campaign, OpenAI says it scanned about 5,034 sensitive locations, or 7% of the codebase, before narrowing down to 2,308 candidates, then confirming 1,724 vulnerabilities in isolated virtual machines. The full series cost $1,213,430 in compute, about $1,279 per exploit. This serves as a reminder that automating security is not free, even when it arrives dressed as a super-assistant.
What limited access changes for defenders
GPT-6 Cyber is currently only available via the API and Codex, and only for a restricted group of Daybreak Red clients. OpenAI plans to gradually expand access to all Daybreak Red clients, while opening up GPT-5.6 Cyber and GPT-6 Astra more widely to other researchers and organizations.
The strategy primarily targets defenders, including operators of critical infrastructures such as power plants or water treatment stations. OpenAI highlights its $1 billion funding program Daybreak for Frontline Defenders and a subgroup called Daybreak Red for Critical Infrastructure, intended for authorized testing. A very concrete limit remains: the more these models become capable of finding flaws, the more their misuse becomes plausible, which the company itself acknowledges.
Key points
- GPT-6 Cyber was presented by OpenAI on November 5, 2026.
- 81 previously unknown critical flaws were found in Firefox.
- 1,724 Linux vulnerabilities were reproduced, 948 exploited.
- OpenAI mentions $1,213,430 in compute costs.
- The model remains reserved for selected Daybreak Red clients.
- Mozilla has already fixed the reported vulnerabilities.
By the numbers
- 99.93% — compliance with safeguards without security instructions, according to OpenAI, 2026.
- 48.5% — score on ExploitGym for GPT-6 Cyber, according to OpenAI, 2026.
- 81 — previously unknown critical vulnerabilities found in Firefox, according to OpenAI and Mozilla, 2026.
- 1,724 — vulnerabilities reproduced in the Linux kernel, according to OpenAI, 2026.
- $1,279 — average cost per exploit on Linux, according to OpenAI, 2026.