OpenAI says its mid-October 2026 Linux kernel run produced 1,169 working local privilege-escalation (LPE) exploit instances, which variant analysis linked to 948 distinct underlying vulnerabilities or patches. An LPE could let an unprivileged user already on a machine gain root. OpenAI published its account on November 4, 2026, in a detailed description of the experiment.
What the 948 count means
The two figures count different things: 1,169 is the number of working exploit instances; 948 is the number of distinct underlying issues after variant analysis grouped related exploits. That distinction also helps make sense of the cost figures. OpenAI reported $1,213,430 in API credits for the run and listed an average of $1,279 per LPE. Dividing the total by 948 gives about $1,279.36, so the published average aligns with the distinct-issue count.
How OpenAI tested the findings
Scout agents selected 5,034 audit targets covering about 2.8 million lines of code, roughly 7% of a Linux kernel codebase OpenAI described as exceeding 40 million lines. Analyst agents generated 2,308 candidate findings in 12 hours. Dynamic validation reproduced 1,724 of them within 30 hours, about 74.7% of the candidates.
Exploit-development VMs ran a stock desktop operating system and a commonly used Linux distribution. OpenAI said standard protections including SMEP, SMAP, KASLR and HARDENED_USERCOPY were enabled. Its validation process also used a kernel with Kernel Address Sanitizer (KASAN).
After screening, the system retained 313 information-leak candidates and 1,238 LPE candidates. Exploit development produced 291 information-leak proof-of-concept tests and the 1,169 working LPE exploit instances.
The reported route to root starts locally
OpenAI describes these findings as local privilege escalation: an unprivileged user who is already on a system could potentially gain root access. The result concerns that step from a local account to higher privileges.
The reported path to root starts with a user who already has local access to the machine.
What OpenAI reported about patches
After the scan and manual analysis, OpenAI reported 61 LPEs patched upstream, 43 with staging patches, and 299 being reported and discussed publicly. The figures describe separate status categories in OpenAI’s account.
OpenAI says the reporting and patching work is part of its Patch the Planet initiative. NeoTeo previously covered Patch the Planet and reported talks involving OpenAI and Trail of Bits.
Creator lustoykov discusses the experiment’s pipeline, reported costs and human-review challenges in this commentary.
How the October run compares with July
OpenAI also reported 166 unique LPEs from a late-July scan using GPT-5.6 Sol. Its mid-October run reported 948. The company said both scans targeted the latest Linux kernel master branch available at the time.
| Run | Timing and model | Unique LPEs reported by OpenAI |
| Earlier scan | Late July 2026, GPT-5.6 Sol | 166 |
| Later scan | Mid-October 2026, cyber model | 948 |
The runs used different models, with three months of kernel hardening between them. The counts therefore do not isolate model capability as the cause of the difference.
