Pre-execution safety oracle for agent actions: submit the tool call you are about to run (shell, http, sql, file, code, env) plus your stated intent, and get a machine-enforceable verdict before you execute it. Decodes what the call does, flags the danger toolkit (rm -rf, reverse shell, curl|sh, SSRF to cloud metadata, credential reads, DROP/DELETE-without-WHERE, path traversal, dynamic eval), and binds it to your intent (allowedHosts/allowedPaths/readOnly/noNetwork) - only a fully pinned, clean, intent-matched call is auto-exec-safe. Hybrid: a deterministic, uninjectable detector engine (authoritative) plus an LLM classifier that can only raise the risk. Fails closed. Detection of known-dangerous patterns, not a proof of safety; it never executes the call.