OpenAI said on 4 November 2026 that an automated pipeline running an upcoming, unreleased cyber model found 948 distinct Linux kernel vulnerabilities that let an ordinary local user gain root, built a working exploit for them and wrote a candidate patch for each. Its Cyber team ran it in mid-October against the latest kernel, on stock desktop systems from a common distribution with standard protections such as SMEP, SMAP and KASLR enabled.
The numbers, as reported by Help Net Security and Unite.AI: agents narrowed the audit to about 7% of the kernel, reported 2,308 possible bugs in 12 hours and reproduced 1,724 of them in isolated virtual machines over 30 hours. The exploit stage produced 291 information-leak proofs of concept and 1,169 working root exploits, which grouped into 948 underlying flaws. The run cost $1,213,430 in API credits, about $1,279 per exploit. OpenAI says 10 exploitation techniques in a 166-exploit sample were new, and withheld their details.
Patches are not all shipped. After about 10 days, 61 of the bugs were patched upstream, 43 had staging patches and 299 were being reported and discussed publicly. OpenAI says it has sent hundreds of reports and patches, starting with the most serious at the kernel community's suggestion, and will send the rest with its Patch the Planet partners. Unite.AI stresses that 948 "is not a count of merged patches".
Limits OpenAI admits: in an early run, 20 of 200 claimed exploits reached root through a different zero-day than the assigned one, so it added an independent check; early patches sometimes had unnecessary code. Humans review every exploit, patch and severity rating before it goes to maintainers. OpenAI's advice to defenders: treat an unpatched low-privilege account as effectively root for a well-resourced attacker, and add layers such as sandboxes and isolation (Firecracker, gVisor).
