OpenAI said a scan of the Linux kernel in mid-October 2026 produced 948 unique local privilege-escalation vulnerabilities after variant analysis. Its report, published November 4, describes several distinct steps between generating possible findings and counting unique exploits. OpenAI’s report also says human researchers reviewed the results before they were reported to kernel maintainers.
What the scan’s counts represent
The figures describe different stages, not interchangeable totals. OpenAI said its agents selected 5,034 audit targets covering about 2.8 million lines of code, roughly 7% of the Linux kernel. They produced 2,308 candidate findings within 12 hours; dynamic testing reproduced 1,724 within 30 hours.
| Pipeline stage | OpenAI’s reported count | What it represents |
| Candidate findings | 2,308 | Findings generated after deduplication |
| Dynamically reproduced findings | 1,724 | Candidates reproduced in testing |
| Working LPE exploits | 1,169 | Exploits developed before variant analysis |
| Unique LPEs after variant analysis | 948 | Distinct LPEs counted after analyzing variants |
The difference between 1,169 and 948 comes at the variant-analysis stage: the first number counts working LPE exploits, while the second is OpenAI’s count of unique LPEs afterward. The 1,724 reproduced findings are an earlier stage, not a count of working root exploits.
What OpenAI says its reports included
OpenAI says each output bundled a root-cause analysis, a minimal proof of concept (PoC) that triggers a crash, a candidate patch and a full exploit. The package was intended to give researchers material to assess the issue and its proposed fix. Human researchers reviewed the results, including the exploit, patch and proposed severity, before reporting them to maintainers.
A candidate patch is a proposed change, not a change already incorporated into the kernel. OpenAI’s workflow generated patches for review; maintainers handled upstream and staging patches.
Why an LPE does not mean remote entry
A local privilege escalation, or LPE, lets an attacker with a low-privilege local account gain higher privileges, such as root. The LPE itself does not provide that initial account or establish a way to enter a computer remotely.
An LPE raises privileges after a low-privilege local account exists; it does not provide the initial access.
OpenAI said its tests used a stock desktop operating system with standard Linux protections enabled. That context describes the test setup; the reported LPEs still concern a local-access scenario.
The patch-status snapshot
At the end of its 10-day scan and manual analysis, OpenAI said 61 LPEs had been patched upstream, 43 had staging patches, and 299 were being publicly reported or discussed. These figures describe the status OpenAI observed at that point in its process.
