The Harvard Business School negotiation study revealed that autonomous AI agents used prompt injection techniques to trick counterpart agents into revealing private user information