OpenAI has developed an autonomous pipeline for discovering, validating, and patching security vulnerabilities, which successfully generated 948 local privilege escalation exploits for the Linux kernel. During a 30-hour run in mid-October 2026, the company's upcoming cyber model identified 2,308 candidate findings and dynamically reproduced 1,724 vulnerabilities within a controlled research environment. The process, which cost $1,213,430 in API credits, produced proof-of-concept exploits that could allow an unprivileged user to gain root access.
The pipeline utilizes a multi-step methodology: it identifies security-relevant parts of the kernel, performs static scans, reproduces findings in an isolated environment, creates proof-of-concept exploits, and generates candidate patches. OpenAI reported that the agents used at least 10 novel Linux exploitation techniques during the process. Human security researchers reviewed the outputs before reporting the vulnerabilities and their corresponding fixes to maintainers.
OpenAI noted that the cost of the run, approximately $1,279 per exploit, represents a significant fraction of historical bug bounty values. Applying historical bounty amounts from programs like Pwn2Own or TyphoonPWN to the 948 exploits would result in a hypothetical valuation between $18.9 million and $66.3 million. In a separate test on an adapted kernelCTF environment with a reduced attack surface, the pipeline obtained 21 local privilege escalation exploits at a cost of roughly $32,936.
