OpenAI cyber model exploits 948 Linux kernel bugs
OpenAI says an unreleased cyber model found and exploited 948 Linux kernel privilege-escalation bugs for about $1.2 million in API costs, at roughly $1,279 per working root exploit.
OpenAI reported that its automated pipeline, running on an upcoming cyber model, produced 948 working local privilege escalation (LPE) exploits against the Linux kernel. Each exploit lets an unprivileged local user gain root. The pipeline also wrote candidate patches for the bugs.
The run took place in mid-October 2026 and cost $1,213,430 in API credits. The pipeline picks out security-relevant parts of the kernel, scans them statically, reproduces candidate bugs in isolated VMs, writes proof-of-concept exploits and drafts patches. Human researchers then review the results. Of 2,308 candidate findings, 1,724 were reproduced. The pipeline produced 291 information-leak PoCs on top of the 948 LPEs. The targets were stock desktop installs of a commonly used Linux distribution, tested under the same threat model as Pwn2Own-style competitions. OpenAI did not name the distribution.
On an adapted kernelCTF environment, which has a reduced attack surface, the pipeline produced 21 LPE exploits for about $32,936. By OpenAI's count, kernelCTF has accepted 26 confirmed 0-day LPEs so far in 2026. OpenAI also says the agents used at least 10 exploitation techniques it believes are new.
The company says the new model found more than five times as many novel LPEs as a GPT-5.6 Sol run in July, even though the kernel had three months of hardening between the two scans. This builds on June's Patch the Planet work, when GPT-5.5 Cyber produced 24 LPE exploits.
OpenAI says it is reporting critical bugs first, following guidance from kernel maintainers. It has not said how many patches have been accepted upstream. During the scan, other contributors patched more than 100 Linux LPEs. OpenAI argues that for now, a low-privileged account on an unpatched system should be treated as equivalent to root access for a well-resourced attacker.