FLASH CRASHES IN MULTI-AGENT SYSTEMS USING MINORITY GAMES AND REINFORCEMENT LEARNING TO TEST AI SAFETY