OpenAI says its mid-October 2026 Linux kernel campaign produced 948 distinct local privilege-escalation findings after variant analysis. Local privilege escalation (LPE) means a person or process with low-level access to a machine gains higher privileges; the figure is not a count of remote attacks. OpenAI’s account describes several different counts from the campaign, each tied to a separate stage or result.
The 948 figure counts distinct local findings after variant analysis, not remote entry points.
What the 948 figure counts
OpenAI reported 2,308 candidate findings, 1,724 vulnerabilities reproduced dynamically, and 1,169 working LPE exploit results. After analyzing variants, it grouped the vulnerabilities behind those exploit results into 948 distinct findings and patches.
| Campaign stage or result | Count reported by OpenAI | What it represents |
| Candidate findings | 2,308 | Deduplicated candidates produced by the research pipeline |
| Reproduced vulnerabilities | 1,724 | Candidates dynamically reproduced during validation |
| Working LPE exploit results | 1,169 | Successful local privilege-escalation exploit results |
| Distinct LPE findings after variant analysis | 948 | Distinct underlying vulnerabilities and patches behind the exploit results |
| Information-leak proofs of concept | 291 | A separate set of working demonstrations of information leaks |
The 1,169 and 948 counts describe different things: exploit results and distinct underlying findings. The 291 information-leak proofs of concept (PoCs) are another category, not additional LPE vulnerabilities.
The tests involved local access
OpenAI says the main campaign scanned the latest Linux kernel master branch at the time, selecting 5,034 audit targets across about 2.8 million lines of code, or roughly 7% of the kernel codebase. Its reported LPE findings concern a low-privileged user who already has local access. They do not describe remote code execution against arbitrary Linux machines.
OpenAI says it reproduced vulnerabilities in isolated virtual machines with KASAN enabled. Exploit-development tests used a standard-distribution configuration with SMEP, SMAP, KASLR, and HARDENED_USERCOPY protections enabled. These details describe the campaign’s test setup; they do not turn its reported count into a claim that every Linux installation is vulnerable.
What OpenAI reported about patches
At the end of its scan-and-analysis period, OpenAI reported 61 LPEs patched upstream, 43 with staging patches, and 299 being publicly reported and discussed. The figures describe separate reported statuses; OpenAI did not say that all 948 distinct findings had been fixed.
OpenAI also said it had submitted hundreds of reports and patches for critical issues and planned to send remaining findings in stages with Patch the Planet partners.
