Text-Safe Is Not Tool-Safe: The Safety Layer Alignment Skips

A language model that will not write a phishing email will still forward the confidential file, if a document it reads tells it to. The refusal and the action are governed by two different things, and...