OpenAI AI Agents Find 948 Linux Kernel Root Exploits
OpenAI says its AI agents built working root-escalation exploits for 948 unique Linux kernel flaws, at roughly $1.2 million in API costs.
SUMO· the number puzzle
Dialo· learn languages by speaking
ediviewer· read EDI files like documents
giftdetect· the right gift in minutesOpenAI has published the results of an experiment in which a system of AI agents automatically found and exploited vulnerabilities in the Linux kernel. The agents reproduced 1,724 vulnerabilities and developed working exploits for 948 unique bugs that let a local user escalate privileges to root. The run cost about $1.2 million in API spend.
The work was carried out in mid-October 2026 on the then-current kernel version. OpenAI built several types of agents: some selected code reachable by unprivileged users, while others hunted for security bugs, tested whether they were exploitable and generated fixes.
From 5,034 code sections to 948 root exploits
The system first picked 5,034 sections, about 2.8 million lines or roughly 7% of the kernel. Over the next 12 hours the agents flagged 2,308 potential flaws. Each finding was tested with a reproducer in an isolated virtual machine with KASAN enabled, confirming 1,724 of them, or 74.7%.
The agents then produced 291 working information-leak examples and 1,169 exploits for root access. Several exploits targeted the same bug in different ways, so researchers reduced them to 948 unique privilege-escalation vulnerabilities and prepared separate patches. Tests ran on a standard desktop configuration with SMEP, SMAP, KASLR and HARDENED_USERCOPY enabled.
In an adapted Google kernelCTF environment, the agents produced 21 privilege-escalation exploits and 25 information leaks, though SLAB_VIRTUAL had to be disabled because of incompatibility. There is no word yet on whether kernelCTF will accept them.
New techniques, errors and costs
OpenAI says at least 10 previously unknown exploitation techniques emerged from a sample of 166 exploits, covering kernel memory object placement, page preparation and privilege escalation through data modification. The company will not publish details.
Automation was not flawless. In an early run, agents reported 200 successful exploits, but manual review found that 20 gained root through a different bug than the one under test. A dedicated check now confirms the intended flaw is used.
The run cost $1,213,430 in API charges, about $1,279 per unique vulnerability with a working exploit. Roughly $776,000 went to exploit development and validation and about $317,000 to initial code analysis. That figure covers model calls only, not total research costs.
OpenAI staff review findings before submission. Several hundred critical reports with patches have already gone to Linux developers, with the rest to follow, partly with members of the Patch the Planet initiative. The 948 patches are therefore not yet accepted into the kernel. At this scale, validating results and shipping fixes quickly will matter as much as finding bugs.
Apps worth a look
SUMO: the number puzzleTap the right tiles, add them up, hit the target. Easy to learn, hard to master. Free, for all ages.
Dialo: learn languages by speakingTalk to an AI conversation partner; every sentence comes back corrected and explained in your language. 9 languages, A1 to C2.
ediviewer: read EDI files like documentsOpen EDIFACT and X12 messages, see every segment explained, convert to JSON/XML, export PDF and Excel. Runs locally, no upload.
giftdetect: the right gift in minutesPick who it is for, the occasion and their personality. The search engine reads product descriptions from local shops and ranks what fits. Free, no sign-up.
