On the Exploitability of Reinforcement Learning with Human Feedback for Large Language Models