OpenAI cyber model produces 948 working Linux kernel privilege-escalation exploits
Prepared by Sintelligent with the help of SI. Published in SI: . Updated: .
Models can already find and validate vulnerabilities in critical software at scale; defenders gain a fast way to fix them, but the same scale becomes available for attacks.
According to OpenAI's description, a controlled research pipeline built around its as-yet-unreleased cyber model identified 5,034 audit targets covering about 2.8 million lines of Linux kernel code. In 12 hours, analyst agents proposed 2,308 potential findings; dynamic validation reproduced 1,724 vulnerabilities. The agents then created 291 information-leak proofs of concept and 1,169 working local privilege-escalation exploits; after variant analysis, OpenAI counted 948 unique LPEs and prepared patches for the corresponding vulnerabilities. The full run cost $1,213,430 in API credits, or $1,279 per unique LPE. OpenAI estimates that at least 10 new techniques were used in the sampled exploits. Human researchers reviewed the results before submission; hundreds of reports and patches for critical vulnerabilities have already been sent to Linux contributors and Patch the Planet. In one early check, 20 of 200 reported successes used a different vulnerability from the one assigned, after which OpenAI added an independent check of each exploit's path.
Sources 1