Subscribe
0%
Illustration for: OpenAI's agents built 948 Linux kernel root exploits in one $1.2M run. What the numbers show, and what to patch

OpenAI's agents built 948 Linux kernel root exploits in one $1.2M run. What the numbers show, and what to patch

OpenAI says an agent pipeline running an upcoming cyber model turned an October scan of about 7% of the Linux kernel into 948 distinct local privilege-escalation bugs, at roughly $1,279 per working exploit. Ten days later only 61 were patched upstream, so Linux admins should treat any local account on an unpatched box as a path to root.

Listen to this article0:00 / 5:32

OpenAI says a pipeline of AI agents running an upcoming, unreleased cyber model found 948 distinct local privilege-escalation bugs in the Linux kernel in a single October run, and produced 1,169 working root exploits for them before duplicates were merged. The company published the research on November 5 in a post titled "A thousand fewer paths to root", according to Help Net Security, Cyber Security News and SecurityOnline, which all summarized it the same day. OpenAI put the bill at $1,213,430 in API credits, or about $1,279 per exploit, and calls that an upper bound.

Why it matters: kernel root exploits have long been expensive, specialist work. If OpenAI's figures hold, the cost of turning a local foothold into root on Linux just fell by orders of magnitude, and the bottleneck moves to how fast maintainers and admins can ship fixes. Ten days after the run, most of the bugs were still unpatched upstream.

How the run worked

The reports describe a staged pipeline that ran in mid-October against the latest kernel, targeting a standard desktop distribution:

Stage What OpenAI reports
Scout agents Picked 5,034 audit targets, about 2.8 million lines or roughly 7% of the kernel
Analyst agents 2,308 candidate bugs in about 12 hours
Triage agents 1,724 reproduced (74.7%) in about 30 hours, in isolated VMs with KASAN
Exploit agents 1,169 working root exploits and 291 info-leak proofs of concept, tested against SMEP, SMAP, KASLR and HARDENED_USERCOPY
After variant analysis 948 unique bugs, each with a patch written

Cyber Security News reports that about 3.1% of findings were wrong because the model assumed a condition that could not occur. In an earlier run, human reviewers found that 20 of 200 claimed root exploits actually reached root through a different zero-day than the one they were assigned. OpenAI's own conclusion, per Help Net Security, is that even successful agent exploits "need proper validation".

OpenAI also says that in a sample of 166 exploits, the agents used at least 10 exploitation techniques that its researchers, and outside kernel-exploitation experts, considered new. It withheld the technical details.

The cost comparison

OpenAI frames the economics against public bug bounties. Using Pwn2Own's $20,000 and TyphoonPWN's $70,000 payouts, it values the October haul at $18.9 million to $66.3 million. On an adapted version of Google's hardened kernelCTF setup, the pipeline produced 21 root exploits and 25 info leaks over 112 hours for $32,936. OpenAI estimates those would have earned about $1.5 million, and says the pipeline was "net profitable" at the time. These are OpenAI's own valuations, not paid bounties.

The trend line is steep. Per Cyber Security News, a GPT-5.5-Cyber run in June produced 24 Linux privilege-escalation exploits, and a GPT-5.6 Sol run in July found 166 unique bugs for $513,690.

Which model, and who can use it

The research post, as summarized, describes the model only as "upcoming". BERI, writing the same day, ties the work to GPT-6 Cyber, which it says carries a "Critical" cybersecurity rating under OpenAI's preparedness framework. BERI reports it is limited to approved organizations in OpenAI's Daybreak Red tier, with requirements including SSO, MFA, managed devices, FIDO2 keys and IP-allowlisted API access, and no published pricing. BERI also notes that the independent evaluator Irregular found no successful attacks on fully hardened targets, and that neither GPT-6 Cyber nor GPT-6 Astra solved the hardest 15 challenges in its tests.

Patch status

According to Help Net Security and Cyber Security News, ten days after the run:

  • 61 bugs were patched upstream
  • 43 had patches in staging
  • 299 were being reported and discussed publicly

OpenAI says it reported the most serious issues first, as the kernel community suggested, and is sending the rest through partners in its Patch the Planet program. It says security researchers reviewed every exploit, patch and severity rating before anything went to maintainers. No CVE list or affected-subsystem breakdown has been published in the coverage we read.

What to do

OpenAI's blunt framing, quoted by Help Net Security: an unpatched low-privilege account "should be considered functionally equivalent to root access" for a well-resourced attacker. In practice:

  1. Take kernel updates quickly and reboot into them. Hundreds of fixes from this run are still landing; a staged kernel you haven't booted protects nothing.
  2. Treat local access as privileged. Shared hosts, CI runners, shell servers and multi-tenant boxes are most exposed, because the bugs need a local foothold.
  3. Restrict unprivileged user namespaces where your workloads allow it, a common hardening step Cyber Security News recommends.
  4. Isolate untrusted code in a stronger sandbox such as Firecracker or gVisor rather than relying on containers that share the host kernel.

What to watch: how many of the 948 reach stable and LTS kernels, and when distributions ship them. Expect an unusually heavy run of kernel security updates over the coming weeks.

Sources

  1. Help Net Security — OpenAI's upcoming AI model built 948 working Linux root exploits in a single run helpnetsecurity.com
  2. Cyber Security News — OpenAI uncovers 948 Linux kernel vulnerabilities cybersecuritynews.com
  3. SecurityOnline — OpenAI cyber model and Linux kernel vulnerabilities securityonline.info
  4. BERI — GPT-6 Cyber tests found no wins on fully hardened targets beri.net

#OpenAI#GPT-6 Cyber#Linux kernel#privilege escalation#AI agents#vulnerability research#patch now