Summary
OpenAI says an AI-assisted pipeline reproduced 1,724 Linux kernel vulnerabilities and generated 948 distinct local privilege-escalation exploits in a research run. Human researchers reviewed the results, and reporting and patching are under way.
OpenAI says an AI-assisted pipeline reproduced 1,724 Linux kernel vulnerabilities and generated 948 distinct exploits that could let an unprivileged local user gain root access. The company’s mid-October 2026 research run combined automated code analysis, testing in isolated virtual machines, exploitability assessment and candidate patch generation. Human security researchers reviewed the results before reporting vulnerabilities and patches to maintainers.
The work illustrates how AI systems can scale vulnerability research across a large codebase. It also shows why automated findings still need careful validation: a successful exploit must use the vulnerability it was intended to test, and patches require review.
From candidate findings to working exploits
The pipeline first used “Scout” agents to select security-relevant parts of the Linux kernel that could be reached by low-privileged users. It identified 5,034 audit targets, covering about 2.8 million lines of code—roughly 7% of the kernel codebase. “Analyst” agents then produced 2,308 candidate findings.
A candidate report was not counted as a confirmed vulnerability until the system attempted to reproduce it in an isolated virtual machine. The agents used a Linux kernel with Kernel Address Sanitizer, a tool that detects memory errors, and refined proof-of-concept programs against the reported behaviour. OpenAI says this stage reproduced 1,724 findings, or about 74.7% of the candidates, within 30 hours.
The next stage tested whether reproduced issues could enable information disclosure or local privilege escalation. The test machines used a stock desktop operating system and standard protections, including safeguards against certain kernel memory attacks. The system generated 291 proof-of-concept information leaks and 1,169 working local privilege-escalation exploits. After variant analysis, OpenAI counted 948 distinct vulnerabilities behind those exploits, producing 948 distinct patches.
A local privilege-escalation flaw allows someone who already has a local, low-privilege account to gain higher permissions—in this case, root. The findings therefore concern an important boundary within a system, rather than demonstrating remote access to a machine by themselves.
OpenAI reports that the full run cost $1,213,430 in API credits, or $1,279 per distinct local privilege-escalation exploit. The company says these figures reflect the particular models, target configurations and workflow used, and may be reduced through engineering improvements to the pipeline.
Review and patching remain part of the process
The pipeline produced root-cause analyses, proof-of-concept programs, candidate patches and exploits for human assessment. OpenAI says it has sent hundreds of reports and patches for critical vulnerabilities and is working with partners to submit more, following guidance from the Linux kernel community.
The company also describes a failure that shaped its validation process. In an early run, agents reported 200 successful privilege-escalation exploits; human review found that 20 had gained root through a different vulnerability from the one assigned. OpenAI says it subsequently tightened the task constraints and added an independent check that each exploit used its assigned flaw.
Patching was already progressing during the research. OpenAI says that when the scan and manual analysis finished after 10 days, 61 of the local privilege-escalation vulnerabilities had been patched upstream, 43 had staging patches, and 299 were being publicly reported and discussed. Those figures reflect the project’s status at the end of that scan, not a claim that all 948 findings had been fixed.
The run’s significance is not only the number of findings. Automating parts of discovery and testing may help security teams examine complex software at greater scale, while the experience also points to the need for independent checks and timely coordination with maintainers. In OpenAI’s account, automated patch generation is part of that effort, but researchers remain involved in reviewing results and preparing fixes for upstream projects.