In large-scale triage processing scenarios, agents responsible for reading untrusted content are equipped only with read-only tools and no privileges whatsoever, with safety guaranteed through a sandbox isolation mechanism.