OpenAI ran an autonomous pipeline that found and patched 948 Linux privilege-escalation flaws
From the AI digest for Nov 4, 2026
What happened: OpenAI's security researchers built a pipeline that covers the whole vulnerability cycle — discovery, dynamic reproduction, exploit generation and patch drafting — and ran it in mid-October against the Linux kernel with an upcoming cyber model. The run produced 2,308 candidate findings, dynamically reproduced 1,724 vulnerabilities and generated 1,169 working local privilege-escalation exploits — attack paths that let an unprivileged local user gain root — which OpenAI's variant analysis narrowed to 948 distinct vulnerabilities, against a stock desktop Linux distribution, at a cost of $1,213,430 in API credits (about $1,279 per distinct vulnerability by OpenAI's accounting). On a reduced-attack-surface kernelCTF testbed it produced 25 information-leak proofs of concept and 21 privilege-escalation exploits for roughly $32,936. Human researchers reviewed the output, and OpenAI is reporting the vulnerabilities alongside candidate patches with its Patch the Planet partners.
What's genuinely new: June's GPT-5.5 Cyber Linux work — 24 exploits — has become a systematic, priced operation: the mid-October run found more than 5× as many novel exploits as a GPT-5.6 Sol run in July despite three months of kernel hardening, the agents used at least 10 novel exploitation techniques by OpenAI's count, and OpenAI's closing judgment is that a low-privileged account on an unpatched machine should be treated as effectively root access for a well-resourced attacker.
Watch: Whether Linux maintainers accept the candidate patches, whether the exploits are accepted for kernelCTF bounties, the naming and release of the upcoming cyber model, and independent reproduction of the exploit counts and cost figures.
Links: OpenAI · OpenAI on X · Nov 4, 3:00 AM ET