Reinforcement Learning Under Threats