A thousand fewer paths to root
Points and comments are a snapshot, not live.
OpenAI's AI pipeline found 948 Linux kernel local privilege escalation exploits at a cost far below bug bounty rewards.
OpenAI built a five-stage pipeline using an upcoming cyber model to find and exploit Linux kernel vulnerabilities. It identified 5,034 audit targets from 2.8 million lines of code, produced 2,308 candidate findings, dynamically validated 1,724, and generated 948 distinct LPE exploits and patches. The pipeline cost about $1.2 million in API credits, while historical bug bounties for equivalent exploits total $18.9-$66.3 million. The agents also discovered 10 novel exploitation techniques. Results were reviewed by human researchers and reported to kernel maintainers, who had already patched 61 of the vulnerabilities during the 10-day analysis period.
What commenters are saying
An OpenAI team member answered questions, noting stronger models find more vulnerabilities and that they plan to scan Linux again with each new model generation. Commenters expressed both appreciation for the defensive utility and concern about the implications of AI discovering thousands of kernel bugs. Some questioned how many vulnerabilities were truly novel versus duplicates, and whether the Linux kernel's architecture is fundamentally flawed. Others discussed the difficulty of upstreaming so many patches and the need for structural fixes rather than individual patches. The thread also included hiring details and criticism of OpenAI's hybrid work policy.