Despite numerous guardrails, some AI foundation models don’t always seem to like us. Why does this seem to be the case, and what can we do about it?