OpenAI's cyber model pipeline produced 948 working Linux root exploits in one run

OpenAI said on November 4 that security researchers built an autonomous pipeline for discovering, exploiting, and patching vulnerabilities, and ran it once against the Linux kernel using an upcoming cyber-focused model. The pipeline works by finding security-relevant code, scanning it, reproducing candidate bugs in isolated environments, testing whether they are actually exploitable, and generating candidate patches, with human researchers reviewing results before reporting findings and proposed fixes to Linux maintainers.

In that single run, the pipeline reproduced 1,724 known vulnerabilities, produced proofs of concept for 291 information leaks, and validated and generated patches for 948 distinct local privilege-escalation vulnerabilities on a stock Linux desktop configuration with standard protections enabled. Running the full pipeline cost about $1.2 million in API usage, or roughly $1,279 per distinct privilege-escalation bug. OpenAI said the agents discovered at least 10 Linux exploitation techniques it believes are novel.

Based on historical Pwn2Own and TyphoonPwn bounty rates, OpenAI estimated the hypothetical value of the findings at between $18.9 million and $66.3 million. The company said it is working with Linux maintainers and partners in its Patch the Planet program on reports and patches, with the stated goal of helping defenders find and fix bugs before attackers do.

Why it matters

Automating the discovery of hundreds of working root exploits at scale shows how AI could reshape both attack and defense in software security.

Source: OpenAI