It doesn't take much misinformation to compromise a large language model.

Comments