Heidelberg AI News

Video · Fireship ·

OpenAI just found 948 ways to root your Linux box...

In short

The video explains how OpenAI’s GPT-6 Cyber agent swarm reportedly found hundreds of Linux kernel local privilege-escalation exploits and what this means for software security. It argues that increasingly capable cyber models can strengthen defenders while giving attackers an accelerating advantage, making timely patching critical.

Key points

  • The scout narrowed the Linux kernel to about 5,000 reachable audit targets, or 7% of the codebase.
  • The analyst found 2,308 bugs after 12 hours at a token cost of $316,000.
  • The triage agent confirmed 1,724 bugs as real after 30 hours in isolated virtual machines.
  • Exploit agents produced 1,169 working root exploits despite KASLR and SMEP protections.
  • The project cost $1.213 million, or about $1,280 per exploit, versus historical $20,000 bounty payments per Linux LPE.
  • GPT-6 Cyber scored 48.5% on exploitation benchmarks versus 32.1% for GPT-5.6 Cyber.

Written by an AI model from the video’s captions and description.

A model you're not allowed to use just tore through the Linux kernel and came back with 948 root exploits... let's break it down.

Watch on YouTube The opening of the article at youtube.com.

In this story

A tool or model opens in the Heidelberg AI Curriculum: what it does, what it costs, how to start. A repo opens on GitHub.

Related stories

All stories

OfficialOpenAI

Our new blog for science

The OpenAI science blog shares new findings, practical methods, and firsthand accounts from researchers using AI to advance scientific discovery.