OpenAI System Finds 948 Linux Kernel Privilege Escalation Bugs

Analysis agents then produced 2,308 potential findings after deduplication within 12 hours, at an API cost of $316,549. Next came dynamic validation: an agent built a proof of concept and tried to reproduce the issue in an isolated virtual machine running a kernel with Kernel Address Sanitizer, refining its attempt until it worked or ran out of budget. Over 30 hours, 1,724 of the 2,308 candidates (about 74.7%) were reproduced. According to OpenAI, only 3.1% of the remaining cases stemmed from the model assuming an impossible condition. Validation cost $73,330.

Further agents then worked on 313 information-leak candidates and 1,238 LPE candidates in virtual machines with standard Linux protections enabled, including SMEP, SMAP, KASLR and HARDENED_USERCOPY. That yielded 291 information-leak proofs of concept and 1,169 working LPE exploits. After analyzing variants and preparing fixes, researchers counted 948 unique LPE vulnerabilities.

What it does and does not mean

LPE is not remote takeover. An attacker must already be able to run code as a low-privileged user. It remains a serious class of bugs, but it does not mean AI found 948 ways to remotely hijack every Linux machine.

The more striking claim is novelty. OpenAI reviewed a sample of 166 agent-written exploits and says it contained at least 10 new kernel exploitation techniques, which independent kernel security experts also considered new. The company did not publish technical details for safety reasons.

The cost

The full experiment cost $1,213,430 in API credits, or about $1,279 per unique LPE exploit. Exploit development alone took $775,929, of which $691,000 went to LPE candidates. That is far from a casual hobbyist budget, though the figures show how quickly such work is becoming scalable.