RC RANDOM CHAOS

OpenAI's Automated System Finds 948 Linux Kernel Exploits in $1.2M Run

· via Hacker News

Original source

A thousand fewer paths to root

Hacker News →

OpenAI developed an automated pipeline for discovering, exploiting, and patching Linux kernel vulnerabilities using a frontier cyber model. In a mid-October 2026 scan, the system identified 2,308 candidate vulnerabilities, reproduced 1,724 of them in isolated environments, and generated 948 working local privilege escalation exploits—techniques allowing unprivileged users to gain root access. The complete run cost $1,213,430 in API credits, or $1,279 per exploit. On a hardened kernel competition environment (kernelCTF), the pipeline produced 21 LPE exploits and 25 information-leak proofs-of-concept for $32,936.

The five-stage pipeline uses AI to analyze security-relevant code, identify candidates, dynamically reproduce vulnerabilities, create working exploits, and generate patches. Human researchers reviewed all outputs before sharing findings with Linux maintainers. The team discovered at least 10 novel Linux exploitation techniques. The mid-October scan using a more capable cyber model found more than five times as many novel exploits as a July scan with an older model, despite three months of hardening work in between.

OpenAI argues automated patch generation is now essential to keep pace with AI-driven vulnerability discovery. The company expects frontier models will eventually identify most undiscovered vulnerabilities in major software projects. Recommended defenses include defense-in-depth strategies, formal verification, secure-by-default code training, and repeated scanning with each new model generation.

Read the full article

Continue reading at Hacker News →

This is an AI-generated summary. Read the original for the full story.