OpenAI says AI agents found 948 Linux paths to root access in one research run
The company’s cyber pipeline reproduced thousands of kernel flaws and drafted patches for maintainer review, at about $1.2 million in API cost.
OpenAI said Wednesday that artificial-intelligence agents in a controlled research pipeline found and demonstrated 948 distinct ways a low-privilege user on Linux could reach full administrator, or “root,” access.
Linux is the open-source operating-system kernel that powers many phones, cloud servers and supercomputers. In a mid-October 2026 run, OpenAI said its system scanned security-relevant kernel code, reproduced 1,724 validated flaws out of 2,308 candidates, and built working local privilege-escalation proofs for 948 unique issues, along with draft patches for human review.
The company put the API bill at about $1.21 million, or roughly $1,279 per distinct privilege-escalation case. It said Linux maintainers had already patched 61 of the issues upstream by the time analysis finished, with 43 more in staging and hundreds under public discussion. OpenAI is routing reports through its Patch the Planet partners and said human researchers still review exploits and patches before they go to maintainers.
The work stayed inside isolated test machines, OpenAI said, and needed an existing foothold on a system. The company did not publish the new exploit techniques it claimed agents invented.
Source: OpenAI, Nov. 4, 2026