OpenAI's AI agents turned Linux kernel bugs into 948 root exploits
A ten-day, $1.2m automated run found 2,308 candidate kernel flaws, confirmed 1,724 as real, and built 948 working privilege-escalation exploits, most already patched upstream.
OpenAI published results on 4 November from an autonomous pipeline of AI agents built to audit the Linux kernel at scale. Scout agents pick which code to look at, analyst agents read it for flaws, triage agents reproduce candidates in a VM, exploit agents try to turn real bugs into root access, and a final stage drafts patches for human review.
Over a ten-day run in October, the pipeline produced 2,308 candidate findings, reproduced 1,724 of them as genuine vulnerabilities, and built 948 unique local privilege-escalation exploits, bugs that let an unprivileged user become root. The run cost $1.2 million in API credits, using an unreleased OpenAI cyber model. A comparable scan in July, using the older GPT-5.6 Sol, found only 166 exploits for about $514,000. By the time OpenAI published, 61 of the 948 were already patched upstream, 43 had staging patches, and 299 were being discussed publicly through normal kernel disclosure channels. OpenAI says it has seen no evidence of the findings being used in attacks.
Why it matters: this is the same kernel code every Linux server runs, and the rate at which labs can now find and weaponise its bugs just jumped roughly sixfold in three months for similar spend. Expect more unsolicited kernel CVEs arriving via this kind of pipeline, and expect disclosure timelines to get busier.
The caveat: the model used isn’t publicly available, so nobody can reproduce or independently check this run, and OpenAI is both the researcher and the only source for these numbers.