Skip to content
Main Site News Console

Reducing Thousands of Paths to Root

· OpenAI Translated
OpenAILLM

Eliminating a Thousand Paths to Root | OpenAI

November 4, 2026

We developed an automated pipeline for autonomous vulnerability discovery, exploitation, and patch generation. In a single run powered by our cybersecurity model, this pipeline verified 948 local privilege escalation (LPE) vulnerabilities in Linux and generated patches, at an API cost of $1.2 million.

Historically, security research has often required independent and prolonged investigations for each individual finding; today, this process can increasingly be run as a systematic workflow. This post dives into our work in this area.

While vulnerability discovery is becoming increasingly scalable, vulnerability prioritization and verification remain practical bottlenecks. In a controlled research environment, we built a prototype system designed to uncover security flaws in the Linux kernel. The prototype identified and analyzed 2,308 candidate flaws in the Linux kernel, dynamically reproduced 1,724 vulnerabilities, and generated 948 local privilege escalation (LPE) exploits that allow an unprivileged local user to obtain root privileges.

We leveraged this evidence to prioritize vulnerabilities for reporting and developed candidate patches for maintainer review. Guided by input from the Linux kernel community, we began with the most critical vulnerabilities and are currently collaborating with partners in the Patch the Planet initiative to submit additional reports alongside corresponding patches.

In June, as part of our Patch the Planet initiative, we shared how GPT‑5.5 Cyber identified Linux kernel vulnerabilities and developed 24 local privilege escalation exploits to verify and report them. Building on that foundation, we expanded this effort into an end-to-end pipeline powered by our upcoming next-generation cybersecurity model, spanning vulnerability discovery, dynamic verification, exploitability assessment, and patch generation. Following the run, human security researchers reviewed the outputs and reported the discovered vulnerabilities alongside patches.

Summary

Methodology

  • Overall, our pipeline operates as follows: first, it identifies security-relevant components of the Linux kernel; statically scans these audit targets for candidate findings; dynamically reproduces these findings in an isolated environment; builds proof-of-concept (PoC) exploits to demonstrate exploitability; generates patches; and finally, human security researchers review the results.

  • The pipeline’s outputs include root cause analyses, minimal crash PoCs, candidate patches, and full exploits to assist human researchers in assessing vulnerability severity and submitting fixes upstream.

Key Findings

  • In a mid-October 2026 run, the pipeline built on our upcoming cybersecurity model dynamically reproduced 1,724 vulnerabilities and, for a common…, generated proof-of-concept (PoC) exploits for 291 information disclosure vulnerabilities as well as 948 working local privilege escalation (LPE) exploits.