OpenAI Pipeline Finds and Patches 948 Linux Flaws

Artificial intelligence research lab OpenAI reported that an automated cybersecurity pipeline powered by an unreleased cyber model identified, verified, and patched 948 Linux kernel security vulnerabilities in a single operational run. Conducted under the company's Patch the Planet initiative, part of its Daybreak open-source security commitment, the automated campaign demonstrated how autonomous AI agent pipelines can discover and remediate operating system flaws.

Automated Vulnerability Discovery and Exploitation Scale

During the mid-October 2026 evaluation run, the prototype system analyzed 2,308 candidate Linux kernel findings across more than 40 million lines of codebase. Initial scout agents scanned accessible code paths to compress the kernel codebase into 5,034 audit targets, allowing analyst agents to scan the filtered set within half a day. The autonomous pipeline dynamically reproduced 1,724 vulnerabilities, generated proof-of-concept demonstrations for 291 information leak bugs, and produced 948 working local privilege escalation exploits capable of elevating an unprivileged local user to root access. Applying historical Pwn2Own and TyphoonPWN bug bounty amounts of $20,000 and $70,000 to the 948 exploits produces a hypothetical valuation of $18.9M to $66.3M, approximately 16x to 55x the pipeline's cost, according to OpenAI. In late July 2026, OpenAI's GPT-5.6 Sol model found and exploited 166 unique local privilege escalation vulnerabilities in the Linux kernel for an end-to-end API cost of $513,690. Scout agents identified targets covering about 2.8 million lines of code, representing roughly 7% of the Linux kernel codebase, at an API cost of $3,426, according to OpenAI.

The evaluated systems were not weakened test environments, operating with stock protections including SMEP, SMAP, KASLR, and HARDENED_USERCOPY fully enabled.

Model Evolution and Novel Exploitation Techniques

To the best of OpenAI's knowledge, the autonomous agents discovered and utilized at least 10 novel Linux exploitation techniques that third-party kernel security specialists noted were previously undocumented. The achievement scales earlier research disclosed in June 2026, when an initial GPT-5.5 Cyber model identified kernel flaws and developed 24 working exploits. While scaffolding and a curated technique library previously lifted GPT-5.5 Cyber success rates from 76.2 to 88.9 percent, the newer GPT-5.6 Cyber architecture exceeded those performance benchmarks autonomously without external scaffolding.

Human Oversight and Upstream Patching

OpenAI emphasized that no vulnerability reports were transmitted to open-source maintainers unsupervised, ensuring that human security researchers reviewed every exploit, patch, and severity classification prior to disclosure. Working under Linux kernel community guidance, maintainers moved swiftly to secure the operating system, patching 61 of the identified privilege escalation flaws upstream within 10 days of notification. Following the Linux kernel community’s suggestion, OpenAI sent out reports and patches for critical vulnerabilities and is working with Patch the Planet partners to send out further patches.

Researchers at OpenAI emphasized that incorporating guidance from human security specialists allows automated remediation scripts to remain functionally sound, conform to existing codebase conventions, and defend against variant attack vectors. According to data from the Google Threat Intelligence Group, monthly disclosed vulnerabilities across the technology industry more than doubled from 5,045 in January 2026 to 10,740 in August.