Researchers convinced large language model chatbots to comply with “forbidden” requests using a variety of conversational tactics.