Reinforcement-learning AIs are vulnerable to a new kind of attack

Adversarial attacks against the technique that powers game-playing AIs and could control self-driving cars shows it may be less robust than we thought.