New preprint on arxiv. We show how latent adversarial training improves robustness in vision and large language models.