Skip to content
Laksh
← All explainers

AI security

AI just found 948 ways to break Linux

OpenAI pointed a team of AI agents at the Linux kernel. In one run they built working root exploits for 948 bugs. Here's what happened, and what to do tonight.

By 2 min readAI agents and automation

Get the 2-minute PATCH checklistPDF ↓

In short

In November 2026, OpenAI reported that a pipeline of AI agents flagged 2,308 possible Linux kernel bugs, reproduced 1,724 of them, and built working exploits that turn a normal user into root for 948 distinct bugs. The run cost about $1.2M. The bugs are local, and candidate fixes are being reported to the Linux maintainers.

In November 2026, OpenAI published research where a team of AI agents went looking for security bugs in the Linux kernel, the core of the operating system under Android phones, cloud servers and supercomputers.

What the agents did

The pipeline used several kinds of agents, powered by an unreleased model OpenAI calls "an upcoming cyber model". Some scouted the code for possible bugs, some analysed them, and others tried to prove each one by actually triggering it in an isolated virtual machine.

The funnel looked like this:

Stage Count
Possible bugs flagged 2,308
Reproduced as real 1,724
Distinct bugs with a working root exploit 948

Each of those exploits turns a normal user account into full admin (root). The agents produced 1,169 working exploits in total; 948 is the number of distinct bugs behind them.

The cost

The whole run cost $1,213,430 in API credits, which OpenAI calls more of an upper bound. Using historical bug bounty prices from Pwn2Own and TyphoonPWN, OpenAI puts a hypothetical value on the finds of $18.9M to $66.3M.

The good news

The same pipeline wrote candidate fixes. Human researchers reviewed the results and are reporting the bugs, with patches, to the Linux maintainers. Not all of them have shipped yet.

What this does and doesn't mean

  • These are local bugs. An attacker needs a normal account, or code already running on the machine, first. This isn't a remote, zero-click hack.
  • The exploits were tested on a common desktop Linux distribution. OpenAI didn't test them on Android phones.

The tile walls and funnel in the video are simplified illustrations. The text cards are exact quotes from OpenAI's post.

The warning worth remembering

OpenAI's own conclusion is blunt: on a device that isn't up to date, a normal account should be treated as functionally equivalent to root access for a well-resourced attacker. AI finds bugs faster now, and a fix only protects you once it's installed.

So turn on automatic updates on your phone and laptop tonight, and restart when they ask. The checklist above walks through it for Android, iPhone, Windows, Mac and Linux.

Sources

  1. 01
  2. 02
  3. 03

Spotted a mistake? Tell me and I'll fix it.

Newsletter

New explainer? You hear first.

Each one is a single AI concept, explained in plain words, with the sources linked.

One email when a new explainer is out. No spam.

Watch next