OpenAI's AI found 948 ways to seize root on Linux

OpenAI's autonomous security pipeline scanned the Linux kernel and came back with 948 distinct working local privilege-escalation exploits — ways for an ordinary logged-in user to grab root access — each paired with a candidate patch drafted by the same system.

The run started from 2,308 candidate findings, reproduced 1,724 as real vulnerabilities, and cost $1,213,430 in API credits, about $1,279 per exploit; within 10 days of the mid-October scan, 61 flaws were already patched upstream and 43 more had staging patches.

The twist for defenders and attackers alike: sampling the exploits turned up at least 10 exploitation techniques that outside kernel experts had never seen — evidence, OpenAI says, that stronger models don't just find more bugs, they invent new kinds of break-ins.