OpenAI reveals autonomous vulnerability pipeline: a run that produced 948 local privilege escalation vulnerabilities in Linux

According to disclosures from OpenAI's official website, the company has established an autonomous security research pipeline covering vulnerability discovery, dynamic verification, exploitability assessment, and patch generation. In a run centered on an upcoming network model, the pipeline completed scanning and verification of targets related to the Linux kernel, generating 948 exploits that can achieve local privilege escalation (LPE) on commonly used Linux distributions, with an API cost of approximately $1.213 million, or about $1,279 per LPE exploit.

It was reported that the pipeline identified and analyzed a total of 2,308 candidate discoveries, dynamically reproducing 1,724 vulnerabilities, and producing 291 information leakage proofs of concept (PoC); the target virtual machine ran a standard desktop system with complete protection. The Scout agent initially identified about 5,034 audit targets (approximately 2.8 million lines of code, accounting for about 7% of the kernel), and the Analyst agent produced candidate discoveries in about 12 hours. OpenAI stated that it has prioritized reporting critical vulnerability patches as recommended by the Linux community and continues to advance the remaining reports in collaboration with Patch the Planet partners.

OpenAI also mentioned that in the adapted kernelCTF environment, an additional 25 information leakage PoCs and 21 LPEs were obtained; among 166 sampled exploits, the agent reportedly used at least 10 newer Linux exploitation techniques that it deemed more recent. The company emphasized that maintainers are quick to fix: over 100 LPEs have been patched or entered temporary patches within about 10 days of scanning, and approximately 300 others have entered public discussion.