# OpenAI detects 948 Linux flaws that allow administrator privileges to be obtained

> OpenAI uses AI to detect 948 Linux flaws that allow users with local access to obtain administrator permissions and proposes patches
> By Covadonga Fernandez · Artificial Intelligence · 2026-11-05

OpenAI claims to have identified 948 flaws in the Linux kernel that allow a user with local, unprivileged access to obtain administrator permissions. The company used artificial intelligence agents to locate the flaws, demonstrate their exploitation and generate proposed fixes, which were subsequently reviewed by human researchers. The work, published on November 4, cost $1.21 million in API usage.

The work shows how AI can help detect security flaws in the software used by Android devices, cloud services and supercomputers. The challenge is turning those findings into reliable fixes that reach users.

## Division of tasks among agents

The agents divided up the tasks. Some selected the parts of the code to be reviewed; others looked for vulnerabilities. The system then tried to reproduce them on isolated machines, verify their impact and prepare the patches.

Although the kernel exceeds 40 million lines, the detailed audit focused on about 2.8 million, roughly 7% of the code. The agents prioritized components relevant to security and accessible from accounts with few permissions. The 948 cases correspond to local privilege escalations. This means the attacker needs prior access to the system as an unprivileged user to attempt to gain administrative control. They are therefore not 948 remote entry points into any machine running Linux.

**[OpenAI buys Pierre Computer so Git can keep up with agents](https://observatorioblockchain.com/en/ai-agents/openai-compra-pierre-computer-codex-git-agentes/)**

Automatic patch generation is a central part of the experiment, but the patches must pass review by the project maintainers. Their creation does not mean they are accepted or available in the updates users receive.

## $1,279 in AI usage per bug

[OpenAI](https://openai.com/index/a-thousand-fewer-paths-to-root/?utm_source=chatgpt.com) acknowledges that the first proposals could solve the exploit test and still be incomplete or introduce other problems. The company also found that, in an initial test, some agents obtained administrator permissions by exploiting a different flaw from the one they were supposed to investigate. This forced it to tighten controls.

The company spent an average of $1,279 on AI usage for each bug that allowed administrator permissions to be obtained. That figure does not include the researchers' work or the other technical resources used. The results show that AI can make finding and fixing bugs easier, although those same capabilities can also be used to prepare attacks.
