Stop talking about “rogue” agents and start treating every tool call as a…

Stop talking about “rogue” agents and start treating every tool call as a runtime principal you can control. The recent wave of AI-hacking stories points to ordinary engineering failures: weak containment, overly broad tool permissions, and blind monitoring. The useful move is procedural and structural: make every tool invocation cross a boundary you can authorize, observe, rate-limit, and kill. Treat tool calls like networked services—add explicit auth, telemetry, quotas, and a kill path before you hand them any side effects.