OpenAI security researchers developed an autonomous system to find and fix software flaws that successfully generated patches for 948 local privilege escalation vulnerabilities in the Linux kernel. During a mid-October test using an upcoming cyber model, the pipeline reproduced 1,724 vulnerabilities and created proofs of concept for 291 information leaks. The campaign targeted a stock Linux desktop configuration with standard security protections and cost OpenAI $1,213,430 in API credits, or about $1,279 per patched vulnerability.
The system’s performance marks a jump from a July test where GPT-5.6 Sol found 166 vulnerabilities for roughly $500,000. OpenAI noted that the agents used at least 10 novel Linux exploitation techniques and estimated that the 948 exploits found in the latest run would have a hypothetical bounty value between $18.9M and $66.3M based on rates from competitions like Pwn2Own. In further testing on a reduced-attack-surface kernelCTF environment, the pipeline identified 21 privilege escalation exploits and 25 information leak proofs of concept.
