OpenAI finds 948 Linux vulnerabilities via AI cyber model

OpenAI announced November 4 that its autonomous security pipeline discovered and generated patches for 948 local privilege escalation vulnerabilities in the Linux kernel during a single mid-October test run. The breakthrough involved a prototype using an unreleased cyber model that validated 1,724 vulnerabilities from 2,308 candidates and created working root exploits for each flaw, all at a cost of $1.2 million in API credits. The research also uncovered at least 10 novel Linux exploitation techniques previously unknown to the security community.

OpenAI's autonomous vulnerability discovery pipeline achieved a major security breakthrough by identifying and generating patches for 948 distinct local privilege escalation (LPE) vulnerabilities in the Linux kernel. The pipeline, powered by an upcoming cyber-focused AI model, scanned the Linux kernel-containing over 40 million lines of code underpinning Android devices, cloud infrastructure, and supercomputers-and narrowed its focus to security-relevant areas accessible to unprivileged users.

The process identified 2,308 candidate Linux kernel findings, dynamically reproduced 1,724 vulnerabilities in approximately 30 hours, and ultimately produced 948 working local privilege escalation exploits capable of allowing unprivileged local users to gain root access. Each vulnerability came with a root-cause analysis and candidate patch for maintainer review. The entire campaign cost $1,213,430 in API credits, equating to roughly $1,279 per unique flaw-a dramatic reduction compared to historical human bounty costs for Linux exploits.

The research represents a substantial improvement over OpenAI's previous effort in July 2026, when an earlier cyber model discovered 166 unique LPEs for $513,690 on a less hardened kernel. Within 10 days of the mid-October run, OpenAI noted that over 100 of the discovered LPEs had already been patched or had staging patches from other contributors, with approximately 300 more being publicly discussed, demonstrating the severity and relevance of the findings.

A sample of 166 exploits contained exploitation techniques that both OpenAI's team and third-party kernel experts believed were novel, spanning persistent elastic objects, page grooming, data-only privilege escalation, and arbitrary type confusion. The 948 figure reflects the number of distinct patches generated after variant analysis, rather than a count of merged patches. OpenAI plans to distribute the remaining vulnerability reports and patches in stages through its Patch the Planet initiative, a collaborative program launched in June 2026 with Trail of Bits, HackerOne, and other partners that pairs automated AI discovery with human security engineer review.