AI agents found 948 ways to become admin on Linux

OpenAI pointed AI agents at the Linux kernel, the software at the heart of most computers and servers. The agents flagged 2,308 possible security holes, proved 1,724 were real, and wrote working attacks for 948 of them. Each one is a way for someone with limited access to take full control of a machine.

But a human still holds the keys.

Linux teams had already closed 61 of the holes before the research wrapped, and every fix the agents wrote still needs an expert to check it. Bug-hunting that once took specialist teams weeks now runs overnight as a paid job. The machine finds the doors. People decide which ones get locked.

The journey resets tomorrow.

A new story, a new road, every day. That is the point of coming back.