Linux admins have a new reason to patch and reboot promptly. OpenAI says a single October run of a cyber model it has not released yet found 948 separate kernel flaws, each of which could let a normal local user become root.
The price tag is as notable as the count: about $1.2 million in API credits, which OpenAI itself treats as a ceiling. It also says it wrote a fix for every bug, and that people reviewed the findings before anything was reported to maintainers.
OpenAI’s research write-up, A thousand fewer paths to root, was unreadable from our side because openai.com refused the request, so the figures here are the ones Cyber Security News relayed from it. Maintainers are already working through the reports. Ten days in, 61 of the bugs had landed upstream, 43 more had fixes lined up and 299 were being debated in public.
How it worked: agents first marked out about 7% of the kernel, some 2.8 million lines reachable by a low-privileged user, and raised 2,308 suspects in half a day. Further agents wrote test code and ran it in virtual machines with the Kernel Address Sanitizer on, which confirmed 1,724 of them. Exploit agents then attacked kernels with SMEP, SMAP, KASLR and HARDENED_USERCOPY enabled, ending with 1,169 working privilege-escalation exploits. Once variants of one flaw were merged, 948 distinct bugs remained.
OpenAI also priced the result against bug bounties. At Pwn2Own’s $20,000 or TyphoonPWN’s $70,000 a bug, the set would be worth $18.9 million to $66.3 million, against its own cost of roughly $1,279 per exploit. On a hardened kernelCTF target, the same method produced 21 exploits for about $32,936.
This builds on the June Patch the Planet programme, where GPT-5.5-Cyber wrote 24 Linux exploits. A July run found 166 bugs, so the jump to 948 came despite a quarter of extra kernel hardening in between.
OpenAI is keeping the details of at least 10 new exploitation techniques private. It also flagged a quirk from an early run: 20 of 200 apparent successes gained root through a different, unassigned zero-day, which is why a person has to check how each result was reached.
The advice for anyone running Linux servers is to treat a low-privileged login on an unpatched machine as root in the hands of a well-funded attacker. Apply kernel updates, restart into the new kernel, and restrict unprivileged user namespaces if you can. Hundreds of reports have gone out, and the rest will follow through Patch the Planet partners under OpenAI’s Daybreak programme.
