Adversarially trained neural representations may already be as robust as corresponding biological neural representations.