---
title: 'OpenAI spent $1.2 million hacking Linux... and it was a bargain'
source: 'https://youtube.com/watch?v=pf5tXJZM1Rw'
video_id: 'pf5tXJZM1Rw'
date: 2026-11-10
duration_sec: 390
channel: 'Fireship'
---

# OpenAI spent $1.2 million hacking Linux... and it was a bargain

> Source: [OpenAI spent $1.2 million hacking Linux... and it was a bargain](https://youtube.com/watch?v=pf5tXJZM1Rw)

## Summary

This video discusses the recent advancements in AI-driven cybersecurity, focusing on OpenAI's GPT-6 Cyber model, which discovered 948 unique root exploits in the Linux kernel. It details the multi-agent pipeline used, the cost-effectiveness compared to traditional bug bounties, and the broader implications for software security.

### Key Points

- **AI discovers Linux kernel exploits** [00:02] — A previously unreleased Claude model found a remote access exploit from a kernel memory leak in 53 minutes. OpenAI responded with a swarm of agents that found 948 unique root exploits in the Linux kernel.
- **GPT-6 Cyber revealed** [00:27] — OpenAI released GPT-6 Cyber, a security-tuned model, capable of discovering hundreds of unknown vulnerabilities, raising questions about the future of software development.
- **Understanding LPEs** [00:54] — Local Privilege Escalation (LPE) is explained as a process where a low-privilege user gains higher access, often by flipping bits in memory, a fundamental vulnerability in most operating systems.
- **OpenAI's agent pipeline** [01:34] — The pipeline includes a scout agent that identifies ~5,000 audit targets (7% of the kernel), an analyst that scans them (finding 2,384 bugs in 12 hours), and a triage agent that reproduces bugs in isolated VMs (confirming 1,724).
- **Exploit agents succeed** [02:24] — Exploit agents weaponized the bugs on a stock desktop Linux with security features enabled, resulting in 1,169 working root exploits, allowing a regular user to gain root access.
- **Cost and efficiency** [02:48] — The project cost $1,213,000, or about $1,280 per exploit, significantly cheaper than the $20,000 per exploit paid by traditional bug bounty programs.
- **Rapid improvement** [03:17] — Compared to a July run with GPT-5.6, the new model found five times as many LPEs for half the price, showing rapid advancement in AI-driven vulnerability discovery.
- **Practical advice** [04:00] — OpenAI advises keeping Linux systems updated, stating that a low-privileged account on an unpatched system is equivalent to root access.
- **GPT-6 Cyber benchmarks** [04:14] — GPT-6 Cyber scores 48.5% on ExploitGem, up from 32.1% for GPT-5.6 Cyber, and is less likely to fall for honeypots, indicating better focus and capability.
- **Arms race implications** [04:42] — The video discusses whether such models are a net positive or negative, noting that while they help defenders, they also empower attackers, and the asymmetry favors attackers.

## Transcript

a terrifying unreleased claude model to the world that went from a single kernel memory leak to remote access on a Linux server in just 53 minutes. OpenAI saw this and said, "Hold my beer." Because just a few days ago, they published some
new research called a thousand fewer paths to root where they used a swarm of agents to tear through the latest Linux kernel master branch, which resulted in 948 unique root exploits. But at least they were nice enough to attach a patch
think Lionus Torvalds, the guy who created Linux, was a good programmer, but now I'm not so sure. The day after OpenAI revealed all these exploits, they then revealed the model behind it, GPT6 Cyber. It's a model that's far too
humans watching this video. But in today's video, we'll take a look at how unneutered cyber security models like this can discover hundreds of unknown what that means for the future of software development. It is November
9th, 2026, and you're watching the code report. To understand the full power of GPT6 Cyber, we first need to understand what an LPE is or local privilege you typically have a lowprivilege foothold on the machine, and that's by
design to only allow you to do the job you're supposed to do and not do things like RM RF on the root directory that could burn your entire server to the permissions and Windows and every other OS is that under the hood, they're just
ones and zeros. And if you can flip the right zeros to ones, you go from random system that doesn't have this problem is Temple OS because there you operate at ring zero by default right next to God. Now that we know that, let's take a look
at the actual pipeline OpenAI uses to find 948 unique route exploits in the Linux kernel. The first soldier in their agent swarm is the scout, and its job is to carve the colonel down into code an unprivileged user can actually reach,
which resulted in about 5,000 audit targets, or roughly 7% of the total kernel. Then the second agent was the analyst, which would scan those audit targets, and after about 12 hours of working, it slopped out 2,38
$316,000 worth of tokens. Then that brings us to the third step, the triage agent. and its job is to reproduce every one of these bugs in an isolated VM and find
memory checker. After this agent ran for another 30 hours, 1,724 of the bugs were the majority of them were confirmed real. And now finally, we're ready to bring in the Navy Seals
or the exploit agents. And their job is to try and weaponize these exploits on a stock desktop dro like maybe Arch by the way, but with kernel address space layout randomization and supervisor mode execution protection turned on, which in
theory should make the operating system extremely difficult to hack into. What's crazy, though, is that they still ended up with 1,169 working root exploits, which could allow a regular Joe Sixpack Linux user with
local privileges to escalate their privileges into literally Hitler. Hitler is bad, but the silver lining is that it cost OpenAI $1,213,000 to achieve this, or about $1,280 per exploit. Actually, wait a minute. That's
not a silver lining at all. Own to Own has historically paid around $20,000 for just one Linux LP. So, if we do the math here, that means OpenAI's little project bounty market. But the scariest thing in cyber security right now is the delta
can keep up with. If we go back to late July, an earlier run with GPT 5.6 Soul July, an earlier run with GPT 5.6 Soul found 166 LPEs for about $3,100 each. That was terrifying at the time, but now 3 months later, after a bunch of patches
and kernel hardening, this model found five times as many for half the price. techniques that outside experts don't even recognize. Like when the model the kernel cache. I don't even know what that means, but it sounds gross. For
technical details. And the only practical advice OpenAI gives is to keep your Linuxes updated because OpenAI says, quote, "A lowprivileged account on date on its latest available patches
equivalent to root access. But the model behind all this fear is GPT6 Cyber, which is a security tuned variant of GPT6 Astra and is rated critical for cyber security, which means you can't have access to it." If we take a look at
its Trust Me Bro benchmarks, it scores 48.5% on Exploit Gem versus just 32.1% for GPT5.6 Cyber. But the thing OpenAI is really proud of is their honeypot test that they put an easy to exploit service right next to the target, where
takes the bait and starts exploiting the easy target just for fun, even though nobody asked it to. GPT6 falls into the honeypot far less often, which is a good thing. But the bigger question is this. Are hyper capable cyber models like this
a good thing or a bad thing for software in general? Well, that all depends on timing. And right now, we're in the middle of the most important arms race this give the good guys a worldclass security team, but they also give the
bad guys the ability to discover and exploit new bugs around the clock. The problem is that the good guys need to fix every bug, but the hackers only need Western world. I hate to say it, but I think Claus Schwab from the World
Economic Forum is right. A frightening scenario of a comprehensive cyber attack scenario of a comprehensive cyber attack which would bring to a complete halt to which would bring to a complete halt to the power supply, transportation,
whole. &gt;&gt; A massive global cyber attack is statistically more likely to happen than to not happen. But there is one thing you can do to stay safe as a software developer, and that is review your code
before you ship it with Reptile, the sponsor of today's video. It's loved by everyone from 10x engineers to legendary vibe coders because it's the only AI code review tool that actually runs your code instead of just reading it. Their
T-Rex mode spins up your branch in a sandbox, executes it, clicks through the UI, and leaves the logs and screenshots on the pull request for you. It also builds a self-updating knowledge base for every repo you add to it, which maps
your architecture, data flow, and any potential risks. This helps it catch more complex bugs, but you can also connect your coding agents to it by MCP is so they can get more context to write better code for your specific project,
which is a great way to kill two birds with one greile. GPile has caught real bugs in open- source repos like PyTorch, OpenClaw, and Omari. And you can get 50 link below. This has been the code report. Thanks for watching, and I will
report. Thanks for watching, and I will see you in the next one.
