OpenAI describes an automated Linux-kernel security research pipeline that moves from candidate findings to reproduction, exploitability assessment and proposed patches. Its November 4 report places human security review before findings are reported to maintainers.
Three counts describe different stages
The prototype analyzed 2,308 candidate findings and dynamically reproduced 1,724 vulnerabilities. OpenAI says the run produced 948 working local privilege-escalation exploits on a commonly used Linux distribution.
Those are separate stages of evidence. A suspected issue, a reproduced flaw and a demonstrated privilege escalation should not be counted as interchangeable outcomes. Local privilege escalation also describes a particular security boundary; it does not mean every finding provides unauthenticated remote access.
Patching remains a review process
The pipeline generates candidate patches alongside evidence for human researchers. OpenAI says researchers reviewed the results and prioritized reporting, following guidance from the Linux-kernel community.
The report also notes that maintainers and other contributors were already fixing some issues while the scan ran. Generated patches therefore need to be understood as proposals within an active maintenance process, rather than proof that every reported vulnerability is already fixed in every deployed system.
What the cost figure measures
The highlighted run cost about $1.2 million in API credits, according to OpenAI. That measures the reported model usage for the research setup. It is not a complete accounting of every engineering, review or remediation cost.
The practical implication for defenders is a larger validation and patch-review workload alongside faster discovery. Teams evaluating similar systems should track which findings reproduce, which fixes maintainers accept and which deployed versions actually receive them.
This is OpenAI’s account of a controlled research campaign. It does not establish that a particular organization’s Linux systems have been examined or that the generated patches can bypass normal testing and release procedures.