OpenAI published results on November 4, 2026, from a Linux kernel campaign it ran in mid-October. The company reported 948 distinct local privilege-escalation vulnerabilities after analyzing exploit variants, alongside 1,169 working exploits. Those figures count different things: 948 is not the number of working exploits or of fixes merged into Linux. OpenAI’s report
The 948 figure counts distinct vulnerabilities; the reported number of working exploits is 1,169.
What OpenAI’s 948 figure counts
A local privilege-escalation (LPE) vulnerability lets someone who already has local access to a system raise their privileges, potentially to root, Linux’s most powerful user account. An LPE finding by itself does not establish that an attacker can break in remotely.
OpenAI says its pipeline produced 1,169 working LPE exploits, then used variant analysis to group related exploit variants into 948 distinct vulnerabilities. The company also generated candidate patches for human review. A candidate patch is not the same as a fix merged into the Linux kernel.
From candidate findings to distinct vulnerabilities
The mid-October campaign targeted 5,034 areas of the Linux kernel, spanning about 2.8 million lines of code, or roughly 7% of the kernel. OpenAI said the targets were chosen for security relevance and accessibility to low-privileged users.
The counts track separate stages and categories in the pipeline. They are not interchangeable totals.
| Pipeline stage | Reported measure | Count | Timing or scope |
| Candidate generation | Deduplicated candidate findings | 2,308 | Generated within 12 hours |
| Dynamic validation | Findings reproduced in testing | 1,724 | Within 30 hours |
| Initial screening | Information-leak candidates; LPE candidates | 313; 1,238 | Two screening categories |
| Exploitability assessment | Information-leak proof-of-concept programs; working LPE exploits | 291; 1,169 | Later pipeline outputs |
| Variant analysis | Distinct LPE vulnerabilities | 948 | Counted across exploit variants |
OpenAI’s report describes a pipeline that moved from discovery and dynamic validation through exploitability assessment and patch generation, with human review of the results. During validation, the company said most findings that were not reproduced either required elevated privileges or could not be reproduced within the time available; 3.1% reflected incorrect model analysis.
Why human review remained necessary
In an early run, human reviewers found that 20 of 200 exploits marked successful had reached root through a different zero-day vulnerability from the one assigned to them. OpenAI said it later tightened task constraints and added an independent check that each exploit used its assigned vulnerability. That early-run figure describes the reviewed sample, not the full mid-October campaign.
OpenAI also said that, in a sample of 166 exploits, it and outside kernel-exploitation experts it consulted considered 10 techniques novel. The company withheld the full technical details. Its campaign report counted 61 LPEs patched upstream, 43 with staging patches, and 299 publicly reported or discussed at the end of a ten-day scanning and manual-analysis period.
Commentator lustoykov discusses OpenAI’s report, its pipeline and the role of human review. The video is commentary on the company’s account.
What the reported run cost
OpenAI reported spending $1,213,430 in API credits on the campaign. Exploitation assessment accounted for the largest listed stage cost, $775,929.
| Pipeline stage | Reported API-credit cost |
| Scout target selection | $3,426 |
| Analyst candidate generation | $316,549 |
| Dynamic validation | $73,330 |
| Initial screening | $44,194 |
| Exploitation assessment | $775,929 |
| Stated campaign total | $1,213,430 |
The five stage amounts add up to $1,213,428, which is $2 less than OpenAI’s stated total. OpenAI described the campaign costs as an upper bound because it prioritized measuring the methodology and capabilities over optimizing expenses.
