I've seen skills with malicious code slipped into them that quietly take your data. And in agents there are requests that don't get strictly blocked. If you don't put limits on the prompts, the AI will just go ahead and do things.