Artificial intelligence researchers claim to have found an automated, easy way to construct “adversarial attacks” on large language models.