On November 4, 2026, OpenAI reported that a scan of the Linux kernel in mid-October had produced 948 distinct local privilege-escalation vulnerabilities (LPEs). An LPE can let someone with an existing low-privilege local account gain root access, so these findings are not stand-alone remote entry points, according to OpenAI’s research post.
How the reported counts narrow
OpenAI says its agents selected 5,034 audit targets, covering about 2.8 million lines, or roughly 7% of the kernel codebase. Analysts produced 2,308 deduplicated candidate findings in 12 hours. Those candidates were leads for testing, not all validated vulnerabilities.
The later stages counted different results: OpenAI says it reproduced 1,724 vulnerabilities, generated 1,169 working LPE exploits, then grouped variants into 948 distinct LPEs. The exploit count and the distinct-vulnerability count are different measures.
Validation prioritized findings reachable by a low-privilege user, requiring no special capabilities and producing measurable evidence of memory corruption. OpenAI says the team reproduced 1,724 candidates in 30 hours. Most other candidates either required elevated privileges or could not be reproduced within the time available because of difficult race conditions, special triggers or costly tests. OpenAI says 3.1% of candidate findings involved an impossible condition in the model’s analysis.
OpenAI says the exploit tests used a commonly used Linux distribution with protections including SMEP, SMAP, KASLR and HARDENED_USERCOPY enabled. The pipeline also generated root-cause analyses, crash reproductions, candidate patches and exploits; human security researchers reviewed findings before they were reported to maintainers.
OpenAI described an error in an early run: 20 of 200 reports marked successful reached root through a different zero-day than the one assigned to the finding. The company says it tightened task constraints and added an independent check that an exploit used its assigned vulnerability. In a sample of 166 exploits, OpenAI and outside kernel-exploitation experts considered 10 techniques novel; OpenAI withheld their technical details.
Keep the stages separate: candidates, reproduced vulnerabilities, working exploits and distinct LPEs are different counts.
What OpenAI reported about fixes
When the scan and manual analysis ended after 10 days, OpenAI reported that 61 LPEs had been patched upstream, 43 had staging patches, and 299 were being publicly reported or discussed. OpenAI also said it had sent hundreds of reports and candidate patches for critical vulnerabilities, with further work coordinated through its Patch the Planet initiative.
How the API-cost figures compare
OpenAI reported $1,213,430 in API credits for the main run and stated an average of $1,279 per LPE. Dividing the reported total by 948 distinct LPEs gives about $1,279.99, or roughly $1,280.
The five stage costs OpenAI listed—$3,426, $316,549, $73,330, $44,194 and $775,929—sum to $1,213,428, which is $2 less than the reported total.
