OpenAI publishes security research on November 4 finding 948 local privilege escalation flaws in the Linux kernel
The pipeline dynamically reproduced 1,724 of 2,308 candidate issues and produced 948 unique local privilege escalation exploits, with total API cost of about USD 1.213 million.
OpenAI published a security research report on November 4 describing an automated vulnerability discovery, validation, exploit and patch generation system built on a new generation cybersecurity model. In a test against the Linux kernel, the system identified 2308 potential security issues, dynamically reproduced 1724 of them, and produced 948 working local privilege escalation exploits that let low-privileged local users obtain root.
The test ran in mid-October 2026 using a cybersecurity model that has not been formally announced, and covered five stages: discovery, dynamic validation, exploitability assessment, patch generation and human review. In the discovery stage, agents filtered 5034 audit targets spanning about 2.8 million lines of code, roughly 7 percent of the Linux kernel, and the analysis agent produced 2308 deduplicated candidate vulnerabilities within 12 hours at an API cost of 316,549 dollars. In the dynamic validation stage, 1724 candidates reproduced successfully over 30 hours of testing, about 74.7 percent of the total.
The full run cost roughly 1.213 million dollars in API usage, about 1279 dollars per unique privilege escalation flaw. The report also gives a comparison with a July 2026 run using GPT-5.6 Sol, which confirmed 166 unique privilege escalation flaws at an end-to-end cost of about 513,690 dollars; the newer model found more than five times as many novel privilege escalation flaws despite three months of further hardening. In a separate test in the hardened Google kernelCTF environment, the system obtained 25 information disclosure proofs of concept and 21 local privilege escalation exploits over 112 hours at an API cost of about 32,936 dollars. OpenAI said it has already prioritised submitting reports and patches for severe flaws following community advice, and recommends that maintainers rescan existing code as models improve, while using sandboxing and isolation to limit blast radius.