OpenAI says AI agents found 948 Linux kernel flaws

OpenAI has revealed that an unreleased cyber model used in its research autonomously identified and weaponised Linux kernel privilege-escalation flaws, producing 948 distinct patches tied to separate bugs in a single run. The October scanning operation targeted the kernel’s master branch, with AI agents first isolating security-relevant areas accessible to low-privileged users, then flagging 2,308 potential bugs and confirming 1,724 of them.

In total, the experiment generated 1,169 working local privilege escalation exploits and 291 proofs of concept for information leaks, at a reported cost of $1,213,430 in API credits. OpenAI notes that the 948 figure reflects the number of patches created for the bugs behind those exploits, and that the exploits were validated in virtual machines running an unmodified desktop Linux environment with protections such as KASLR, SMEP and SMAP enabled.

OpenAI says kernel exploitation researchers outside its group were not familiar with many of the exploitation techniques uncovered, and full technical details remain private for safety reasons. The company compares the haul against historical bug-bounty payouts, noting that the 948 bugs would have been worth substantially more in traditional bounties.

OpenAI also described a separate, lower-cost run using an adapted Google kernelCTF target, yielding 21 LPE exploits over about 112 hours for roughly $33,000 in API credits. In response, OpenAI says it has submitted hundreds of reports and patches and will roll out fixes in stages through its Patch the Planet initiative, while emphasising layered defences such as sandboxing and isolated runtimes like Firecracker and gVisor.