• Sources: primary
  • Summary: The video summarizes an OpenAI write-up from its Patch the Planet effort with Trail of Bits, in which scout agents survey the kernel for suspicious patterns, analyst agents audit small target areas, and triage agents build proofs of concept and run them in VMs. It cites about 2,300 candidates, 948 working local privilege escalations and 219 working memory leak proofs, at about $1,200 per escalation.
  • Why it matters: It shows a concrete scout, analyst and triage split for running agents over a codebase far larger than any context window.