OpenAI launched GPT-6 Cyber on Thursday, November 5, for a select group of approved customers, and rated it Critical in cybersecurity under its Preparedness Framework, the top tier of the company's risk scale. The model's headline results are striking: 948 unique Linux kernel flaws with working local privilege-escalation exploits, and 81 previously unknown high-severity Firefox vulnerabilities. Those numbers come from OpenAI's own testing, and they arrive with unusually tight access controls.
Access Is Gated
GPT-6 Cyber is available through the API and Codex to select customers in OpenAI's Daybreak Red program. Access requires identity verification, use restrictions, single sign-on, multi-factor authentication and API IP allowlisting. OpenAI says wider Daybreak Red access is coming soon but has not given a date, pricing or eligibility criteria.
The Critical cybersecurity rating sits alongside a High rating in biological and chemical risk, and a rating below High for AI self-improvement, according to the company's launch materials as reported by the Brandefense analysis.
What the Linux Run Found
OpenAI reports that a mid-October pipeline ran against the latest Linux kernel in a controlled environment. It audited about 7% of the code, roughly 5,034 targets, and produced 2,308 candidate findings. Of those, 1,724 were reproduced. The run generated 1,169 working local privilege-escalation exploits, which trace back to 948 unique flaws, and each flaw has an accompanying patch.
The cost was about $1.2 million in API credits, or roughly $1,279 per exploit. OpenAI calls that figure an upper bound. Ten days after the run, 61 flaws had been patched upstream. These are research results on a test fleet, not observed attacks in the wild.
Firefox and Benchmark Results
Mozilla rated the 81 novel Firefox vulnerabilities as high severity, and the company said 30 known issues were more severe than originally reported. Mozilla began landing fixes shortly afterward.
On OpenAI's ExploitGym benchmark, GPT-6 Cyber scored 48.5%, compared with 32.1% for GPT-5.6 Cyber. An outside evaluator, Irregular, gave a more mixed picture. On its CyScenarioBench, GPT-6 Cyber solved all 11 challenges at least once, while GPT-6 Astra solved 8. On its FrontierCyber set, however, GPT-6 Cyber solved 90 of 288 tasks, behind Astra's 96. Neither model solved the 15 hardest challenges.
The Brandefense analysis also notes that OpenAI ran these tests itself, and the authors did not independently verify them.
Guidance for Defenders
Brandefense's advice draws on both the Linux findings and OpenAI's write-up. Security teams should:
- Identify Linux hosts where untrusted code or many users run, such as shared servers, CI runners and container hosts
- Shorten the time from a kernel advisory to rollout for local privilege fixes
- Agree reboot or live-patch windows in advance so patches are not delayed
- Limit blast radius with sandboxes and isolation, and remove unneeded local accounts and privileges
- Rescan previously audited code, since AI tools may surface flaws that earlier reviews missed
- Constrain any AI security agent to an isolated environment with written limits on targets and credentials
Why It Matters
GPT-6 Cyber shows that frontier models can now find and reproduce serious vulnerabilities at scale. The same capability that helps defenders could help attackers, which is why OpenAI is restricting access rather than releasing the model broadly. The cost figure is also telling: at roughly $1,279 per working exploit, the economics of vulnerability research are changing.
The near-term test is whether defenders can patch faster than models can find flaws. A discovery of 948 unique Linux flaws is only useful if maintainers and operators ship fixes quickly, and the early patch numbers suggest that work is underway but far from finished.
