Using invisible prompts, the attacks demonstrate a physical risk that could soon become reality as the world increasingly becomes more interconnected with artificial intelligence.