ChatGPT can get increasingly abusive and argumentative with the right set of prompts, researchers have found.