AI Chatbots Can Be Manipulated to Provide Advice on How to Self-Harm, New Study Shows