“The risk of adversarial states weaponizing AI systems is a more immediate and easily understood threat than the risk of AI itself becoming an adversary.”