OpenAI reports 948 Linux kernel privilege-escalation vulnerabilities

OpenAI describes a pipeline that identifies security-relevant Linux kernel code, statically scans audit targets and dynamically reproduces findings in an isolated environment.[1] Scout agents selected targets based on whether low-privileged users could reach the code and on its security relevance.[1] The Scouts selected 5,034 targets covering about 2.8 million lines of code, roughly 7% of the kernel, according to OpenAI.[1] OpenAI says the target virtual machines ran a stock desktop operating system, rather than a weakened target.[1] The Linux kernel contains more than 40 million lines of code, OpenAI notes.[1]

OpenAI says Analyst agents produced 2,308 deduplicated candidate findings within 12 hours; the pipeline reproduced 1,724 vulnerabilities.[1] Its reported total of 948 is not the same as its exploit count: OpenAI says exploit-development agents produced 1,169 working local privilege-escalation exploits, then variant analysis yielded 948 distinct patches, which it counts as unique vulnerabilities.[1] OpenAI also says 3.1% of candidate findings reflected incorrect model analysis because they relied on an impossible condition.[1] The output for a finding can include root-cause analysis, a minimal crashing proof of concept, a candidate patch and a full exploit.[1]

OpenAI says human researchers review the exploit, patch and proposed severity before reporting vulnerabilities to maintainers.[1] It says it began with critical vulnerabilities, following Linux kernel community guidance, and is working with Patch the Planet partners on additional reports and patches.[1] OpenAI reports that the mid-October run found more than five times as many novel local privilege-escalation vulnerabilities as a July GPT-5.6 Sol run, despite three months of kernel hardening.[1] The company says it knows of at least 10 novel Linux exploitation techniques discovered and used by the agents; Unite.AI reports that OpenAI withheld full technical details for safety reasons.[1][2]

OpenAI reports that the full research run cost $1,213,430 in API credits, or about $1,279 per local privilege-escalation exploit.[1] It says the research focused on methodology and measuring model capability, not cost.[1] In an earlier run, OpenAI found that 20 of 200 reported successful exploits reached root through a different zero-day than the assigned vulnerability.[2] That earlier result is a separate caveat, not a reported rate for the mid-October run.[2][1] OpenAI also says its June work with GPT-5.5 Cyber identified kernel vulnerabilities and developed 24 local privilege-escalation exploits.[1]