# OpenAI Pipeline Produced 948 Linux Kernel Privilege-Escalation Exploits From a Scan of About 7% of the Kernel

The run reportedly cost about $1.2 million in API spend, or roughly $1,279 per flaw, and OpenAI says its agents found at least 10 exploitation techniques it has chosen not to publish.

- Published: 2026-11-09T06:17:39.329Z
- Canonical: https://polylog.news/ai/2026-11-09/openai-pipeline-produced-948-linux-kernel-privilege-escalati
- Publisher: Polylog (AI desk)
- Section: tech
- Sources: [OpenAI](https://openai.com/index/a-thousand-fewer-paths-to-root/), [Anthropic Research (Frontier Red Team)](https://www.anthropic.com/research/team/frontier-red-team)

OpenAI [published a description](https://openai.com/index/a-thousand-fewer-paths-to-root/) on November 4 of a prototype agent pipeline that audited the Linux kernel, reproduced bugs, wrote exploits and drafted patches. In a mid-October run, the system flagged 2,308 candidate findings and dynamically reproduced 1,724 vulnerabilities. It produced working exploits for 948 distinct local privilege escalation (LPE) bugs. [One report](https://startupfortune.com/openais-ai-pipeline-turned-2308-linux-kernel-bugs-into-948-root-exploits/) says the exploit stage generated 1,169 working exploits before variant analysis consolidated them into 948 unique flaws.

The scope was narrow compared with the result. According to [The Hacker News](https://thehackernews.com/2026/11/openai-ai-pipeline-generates-948-linux.html) and [Unite.ai](https://www.unite.ai/openai-pipeline-finds-948-linux-kernel-privilege-escalation-flaws/), scout agents chose 5,034 audit targets covering about 2.8 million lines, roughly 7% of the kernel. Coverage puts the API spend at about $1.2 million. OpenAI says the agents [discovered and used at least 10 novel Linux exploitation techniques](https://interestingengineering.com/ai-robotics/openai-agents-linux-kernel-vulnerabilities-exploits) and that it is withholding the details for safety reasons. [Phoronix](https://www.phoronix.com/news/OpenAI-948-Linux-LPE-Bugs) and others report that upstream maintainers have addressed 61 of the bugs so far and have staging patches for 43 more.

Two limits apply. An LPE requires an existing foothold on the machine, so none of these bugs gives remote compromise on its own. The figures also come from OpenAI's own write-up. Kernel maintainers have confirmed only part of the set through the patches they have merged. Even so, the run is a stronger result than a benchmark score. The exploits ran against real code, and they were reproduced dynamically rather than graded by a model.

Other frontier labs measure the same capability. Anthropic maintains a [Frontier Red Team](https://www.anthropic.com/research/team/frontier-red-team) that studies cyber and other national-security-relevant capabilities in its models. OpenAI's post is the first lab disclosure in this cycle to put a cost per exploitable bug on that work.

## What this means

A cost of about $1,279 per working kernel privilege escalation is far below what researchers usually spend to find and weaponize a bug of this kind by hand. Kernel maintainers, the Linux distributions that backport fixes, and every operator running long-lived kernels in containers and on edge devices are exposed. The limiting factor shifts from finding bugs to fixing them. Merged patches currently cover a small fraction of the 948 flaws. The consensus treats this as a defensive success because OpenAI disclosed the bugs and helped patch them. That view assumes attackers cannot run similar pipelines, and that assumption holds only while the capability stays inside closed labs. The likelier outcome is that comparable pipelines built on open-weight models appear within months. This depends on open-weight coding agents continuing to match the capabilities of closed models. It would be proven wrong if no independent group reproduces kernel exploitation at this scale by mid-2027. Vendors that sell automated patching and runtime kernel protection gain in the meantime.

## What to watch

- The rate at which the remaining flaws receive upstream fixes and identifiers in the Common Vulnerabilities and Exposures (CVE) list. A slow rate means unpatched, known bugs will stay in production kernels for months.
- Whether another lab or an independent security group publishes a comparable kernel run built on open-weight models. That would show the capability is no longer limited to closed labs.
- Whether OpenAI or kernel maintainers publish the 10 withheld exploitation techniques once fixes ship. Publication would let defenders build detections, and it would also give attackers the same information.