Siarhei HolubLinkedInGitHub

Nov 4, 2026, 12:00 AM PT · found Nov 8, 2026, 9:10 PM PT

OpenAI reports its agent pipeline built working exploits for 948 Linux kernel privilege-escalation bugs

Verdict: Vendor-reported · Agents and coding · AI security

OpenAI's post counts 948 distinct local privilege-escalation bugs with working exploits, at about $1.2M. The figures are OpenAI's own, the patches are unreviewed candidates, and no outsider has verified them.

OpenAI says an autonomous pipeline built working exploits for local privilege-escalation bugs in the Linux kernel and counts 948 distinct bugs.

  • Claim: An unreleased cyber model ran against the kernel in October. OpenAI reports 2,308 candidates, 1,724 reproduced, 1,169 working exploits, and 948 distinct bugs after variant analysis. OpenAI
  • Evidence: The run cost about $1.21 million, roughly $1,279 per bug. OpenAI says hundreds of reports and patch candidates have gone to maintainers, and 61 bugs were patched upstream during the 10-day run. SecurityWeek and CSO Online report the same counts. OpenAI, CSO Online
  • Caveats: Every figure is OpenAI’s and no outsider has verified it. The patches are unreviewed candidates, and no kernel maintainer reaction was found. CSO reports an early test where 20 of 200 “successful” exploits used a different bug than assigned, a detail not found in OpenAI’s post. CSO Online
  • Why it matters: If the figures hold, finding exploitable kernel bugs is cheap and maintainer review becomes the limit. SecurityWeek

Sources: OpenAI, A thousand fewer paths to root (primary) · OpenAI news feed (primary) · CSA research note (press) · SecurityWeek report (press) · CSO Online report (press)

Verdict history
  1. Nov 8, 2026, 9:10 PM PT: Pending. OpenAI page failed to open from the research environment. Press summaries conflict on what 948 counts. Resolve by reading the primary.
  2. Nov 8, 2026, 9:40 PM PT: Vendor-reported. The OpenAI page body was still unreadable, but OpenAI's news feed (4 Nov, 08:00 GMT) describes a pipeline that found, validated, and generated patches for 948 local privilege-escalation vulnerabilities. So 948 counts vulnerabilities, each with a patch.
  3. Nov 8, 2026, 9:50 PM PT: Vendor-reported. OpenAI's page was read through our fetch proxy by another session. 948 counts unique bugs with working exploits. The 1,169 exploits and the $1.21M cost are on OpenAI's page, so they are not press-only. CSO Online's wording matches OpenAI, so the earlier note that it contradicted OpenAI was wrong. Calling the bugs patched overstates it, since the patches are candidates for maintainer review. The 20-of-200 early test appears only in CSO. Verdict unchanged because all figures are OpenAI's own.
  4. Nov 9, 2026, 6:40 AM PT: Vendor-reported. Re-checked. The Cloud Security Alliance note of 5 Nov relays OpenAI's figures through trade press, could not read OpenAI's page either, did no verification, and found no maintainer comment on this run. It adds that most bugs were still unfixed upstream about ten days after disclosure and that sources differ on timing. openai.com pages still return 403 on direct fetch. Verdict unchanged.