Linux kernel: OpenAI AI agents exploit 948 flaws to gain root
OpenAI says an AI agent pipeline exploited 948 Linux kernel vulnerabilities to gain root, at about $1,279 each. What it means for your patching strategy.

On 4 November 2026 OpenAI published the research report “A thousand fewer paths to root”: a chain of AI agents, driven by an unreleased cybersecurity model, produced in a single campaign working local privilege escalation (LPE) exploits for 948 distinct vulnerabilities in the Linux kernel. Each exploit takes an unprivileged user to root. The stated cost is $1,213,430 in API credits, roughly $1,279 per vulnerability.
All figures are stated by OpenAI and reported consistently by The Hacker News, Help Net Security, SecurityWeek, CSO Online and ICT Security Magazine: they are not independent measurements. The sources report no CVE identifiers and no exploitation in real attacks, and the technical details of the exploits have not been released.
What happened
The campaign ran in mid-October 2026 against the latest code in the kernel’s main branch. Targets were virtual machines running an unmodified desktop system from a widely used distribution, with standard protections enabled: KASLR, SMEP and SMAP, plus HARDENED_USERCOPY according to ICT Security Magazine. The report, the same outlet notes, names neither the model nor the distribution.
The work extends Patch the Planet, the initiative announced in June in which, according to CSO Online, GPT-5.5 Cyber had developed 24 LPE exploits for the Linux kernel.
How the pipeline works
The stages, described consistently by the sources:
- Target selection: agents isolate the security-relevant parts of the kernel reachable by a low-privileged user, about 7% of the code (5,034 targets, roughly 2.8 million lines, according to ICT Security Magazine).
- Analysis: 2,308 candidate findings in 12 hours.
- Dynamic validation: each finding is reproduced in an isolated virtual machine with the KASAN memory error detector; 1,724 reproduced in 30 hours (74.7%).
- Exploit development: 291 information leak proofs of concept and 1,169 working LPE exploits. Grouping exploits that trace back to the same flaw (variant analysis), OpenAI counts 948 distinct fixes: that is the headline number.
- Patches and human review: agents write a patch for each flaw; researchers check exploit, patch and proposed severity before reporting to maintainers.
According to ICT Security Magazine, exploit development was the most expensive stage ($775,929) and OpenAI considers the total an upper bound. The comparison with the late-July run using GPT-5.6 Sol is stark: 166 LPEs for $513,690 then, more than five times as many now, despite three months of kernel hardening in between. On an adapted version of Google’s kernelCTF, with a hardened kernel and reduced attack surface, the pipeline produced 21 LPE exploits and 25 information leaks in 112 hours, for about $33,000.
OpenAI puts the hypothetical value of the 948 exploits at $18.9 to $66.3 million, using historical Pwn2Own ($20,000) and TyphoonPWN ($70,000) payouts. CSO Online points out that the comparison does not reflect actual payouts.
New techniques and stated limits
- At least 10 previously unseen exploitation techniques, identified in a sample of 166 exploits and new to external kernel exploitation experts as well. According to ICT Security Magazine: two persistent objects for kernel cache spraying, two memory grooming techniques, five data-only techniques and one arbitrary type confusion. Details are withheld for safety reasons. In MITRE ATT&CK terms this is technique T1068, Exploitation for Privilege Escalation.
- The wrong zero-day: in an early run, out of 200 exploits reported as successful, 20 had reached root through a different flaw from the one assigned. OpenAI added an independent check that verifies the path, not just the outcome.
- Imperfect patches: early fixes sometimes added unnecessary code, hurt performance or missed variants of the same flaw (CSO Online, ICT Security Magazine).
- Only ICT Security Magazine reports that, in the kernelCTF campaign, the agents also exploited a QEMU TCG flaw already fixed upstream but not in the version used.
Patch status
After 10 days, once scanning and manual analysis were complete, 61 LPE vulnerabilities had already been fixed upstream, 43 had staging patches and 299 were being reported or publicly discussed. SecurityWeek quotes OpenAI: over 100 had been patched or had staging patches from other contributors. OpenAI has sent hundreds of reports with patches, starting with the most severe at the kernel community’s suggestion; the rest will follow in stages with Patch the Planet partners. CSO Online notes that the report gives no status for all 948.
What it means for European and Italian companies
OpenAI’s warning is explicit: in the near and medium term, a low-privileged account on a system missing the latest patches is effectively root for a well-resourced attacker. A stolen credential, a web shell or a compromised container is therefore worth more than before.
The critical point, highlighted by IDC in CSO Online, is the window between the upstream fix and the distribution kernel actually running in production. For NIS2 entities, which must handle vulnerabilities as part of their risk management measures, kernel patching and scheduled reboots become a governance issue, not just an operational one. The same applies to healthcare and to Linux servers that are hard to take offline.
What to do now
- Update kernels following your distribution’s advisories and schedule reboots: an updated kernel that is not loaded does not protect you.
- Ask your distribution vendors for backport timelines of the fixes.
- Prioritise multi-tenant hosts, CI/CD runners, shared bastions, container platforms and developer workstations.
- Reduce local access: fewer interactive accounts, no unnecessary shells, review of service accounts.
- Isolate untrusted workloads with sandboxes and isolation technologies such as Firecracker and gVisor, as OpenAI recommends.
- Watch for escalation: processes becoming root from unprivileged parents, kernel errors and crashes, unexpected changes to accounts and groups.
- If you use AI agents for vulnerability research, independently verify how each result was obtained.
How ARXDOME helps
ARX Shield shows open vulnerabilities per machine and per package, with their age and those already exploited highlighted, checks hardening against CIS benchmarks and lets you isolate a suspicious machine from the network. ARX Sentry, with its Linux agent, collects logs, security events and the system and software inventory, and groups suspicious executions and account changes: it helps you see which servers still run an old kernel and spot an escalation in progress.
Last updated:


