OpenAI Describes Automated Discovery of 948 LPE Vulnerabilities in the Linux Kernel

OpenAI said its prototype analyzed thousands of findings in the Linux kernel during a testing campaign and, after deduplication, identified 948 unique local privilege escalation vulnerabilities. Complete independent verification of the results is not yet publicly available.

A new OpenAI announcement describes a prototype automated pipeline for the Linux kernel that searched for candidate vulnerabilities, dynamically reproduced them, created functional local privilege escalation exploits, and proposed fixes. After analyzing variants, the company reported 948 distinct fixes, which it uses as the number of unique LPE vulnerabilities.

The testing campaign took place in mid-October, and OpenAI announced it on November 4, 2026. The company also emphasized that this does not constitute a report of active exploitation of these specific 948 vulnerabilities in the wild.

OpenAI Linux LPE Vulnerabilities: What the Company Reports

According to the announcement, the prototype analyzed 2,308 candidate findings in the Linux kernel. The dynamic validation phase reproduced 1,724 vulnerabilities and created 1,169 functional LPE exploits. LPE, or local privilege escalation, can allow an attacker with a regular account to obtain root privileges.

Individual findings do not always represent separate bugs. OpenAI therefore analyzed their variants and created 948 distinct fixes. The company identifies this number as the number of unique LPE vulnerabilities. The difference between 1,169 functional exploits and 948 unique bugs therefore results from deduplication and work with finding variants.

According to OpenAI, the campaign consumed API credits worth $1,213,430. The company says the validation stage alone, during which it reproduced 1,724 findings, took 30 hours. This figure should not be interpreted as the total time for the entire project, since it applies only to the validation phase.

Reports to Maintainers and the Status of Fixes

OpenAI claims it has already sent hundreds of reports and proposed fixes for critical bugs to Linux maintainers. During approximately ten days of processing, the company says 61 LPE bugs were fixed upstream, fixes were being prepared for 43, and 299 findings were publicly discussed.

However, a complete list of CVEs, affected Linux versions, and the resolution status of all findings is not publicly available. It is therefore not possible to determine which distributions or deployments are affected by specific bugs. Users and system administrators should not act based solely on the number 948, but should monitor official security advisories from their distributions and install available updates.

What the Results Mean and What We Do Not Yet Know

If the results are also confirmed in broader practice, they point to a significantly faster automated ability to find and verify local privilege escalations in critical open-source software. Such a pipeline also increases pressure on maintainers, distributions, and operators to review reports and potentially address them with fixes.

Complete independent verification of the number of 948 unique vulnerabilities, the quality of all proposed fixes, or the reproducibility of the exploits is not yet publicly available. Details remain largely private during coordinated disclosure. The announcement is therefore not evidence of a current mass threat to all systems running Linux.

OpenAI compared the hypothetical economic value of the results with historical rewards in bug bounty programs and reported a range of $18.9 million to $66.3 million. This is not the actual market price of the exploits, but rather the company’s modeled comparison.

Further developments will show whether Linux maintainers, distributions, or CVE records publicly confirm specific fixes and affected versions. The possible publication of auditable data about the methodology, duplicate findings, and the acceptance rate of proposed patches will also be important.

Sources

Verified and updated: 11/05/2026 06:31

Sharing