One AI run for $1.2 million: OpenAI showed a system that found 948 ways to take over the Linux kernel

OpenAI published a detailed breakdown of an autonomous pipeline that hunts down, exploits, and patches vulnerabilities in software code.

The biggest number in the post is the price tag. If a customer were billed at API rates for a single full run of the system, the bill would come to $1,213,430.

The post is titled "A Thousand Fewer Paths to Root." Root is the superuser account with full control over a system, and gaining those privileges is the main goal of most attacks on Linux machines.

The system is built as a multi-layered pipeline of AI agents. Some scan code for potential bugs, while others check whether a bug can be turned into a working attack. A third group writes fixes for the confirmed problems.

The Linux kernel served as the testing ground. It runs Android phones, most of the internet servers, cloud infrastructure, and countless embedded devices, so any serious kernel vulnerability can potentially affect billions of systems.

In the first stage, the agents found 2,308 potential vulnerabilities. The system then filtered out the most dangerous ones and ran them on a virtual machine, with the goal of proving that each bug really does allow a full takeover of the system.

In 948 cases, a working exploit confirmed the danger. An exploit is code that abuses a bug in a program to do things a regular user can't. Here, it meant escalating privileges all the way to root.

Then the patch-writing agents stepped in. They prepared all 948 patches but didn't edit the codebase themselves. Every fix was sent to humans for approval first.

Most of the expense comes from tokens. These are the smallest units of text a language model processes, and they're what OpenAI charges for when you work through the API. Each agent generates huge volumes of intermediate text while "thinking," and with thousands of agents running in parallel, costs climb fast.

OpenAI suggests looking at the sum differently. By its math, $1,213,430 works out to about $1,279 per validated exploit. For comparison, this kind of work is usually done by white-hat hackers through bug bounty programs, where a serious vulnerability pays $20,000 to $70,000.

In the blog, OpenAI lays out the calculation:

Applying those historical bounty amounts to the 948 exploits produces a hypothetical valuation of $18.9M–$66.3M, approximately 16x to 55x our pipeline's cost.

By that logic, a million-dollar bill looks like a bargain. The comparison stays theoretical, though – it's a hypothetical valuation, not actual payouts.

OpenAI isn't calling the work entirely new. In June, it ran a similar smaller-scale experiment under an initiative called "Patch the Planet." The latest effort is a much larger follow-up to that approach.

More news
Tags: