OpenAI cyber agents found 948 routes to root in the Linux kernel

OpenAI says an autonomous cyber-security pipeline produced 948 distinct local privilege-escalation cases in the Linux kernel, each backed by a working exploit and a candidate patch. Such flaws can let an unprivileged local user gain root access on a vulnerable system.

The prototype used an upcoming cyber model across discovery, dynamic validation, exploitability testing and patch generation. It analyzed 2,308 candidate findings, reproduced 1,724 vulnerabilities in isolated environments and built 1,169 working privilege-escalation exploits. Variant analysis reduced those results to 948 unique cases and patches.

The full run cost about 1.21 million dollars in API credits. OpenAI says 80% of known CVEs in the evaluated set were found in under 30 minutes each. The company also reports that sampled exploits used several techniques that outside kernel experts considered new.

The result does not mean 948 new CVEs have already been accepted or fixed. Human researchers reviewed outputs before disclosure, and Linux maintainers were already patching some issues while others remained under discussion. OpenAI says it has sent hundreds of reports and patches for critical vulnerabilities and is working with partners on the rest.

The practical shift is that vulnerability discovery and exploit validation can now be run at a scale that makes human triage, responsible disclosure and maintainer capacity the main bottlenecks.