They've added a lot of safety features. Knowing very little about LLMs, I would assume these prepended prompts are using up a chunk of the limited "attention" the transformer has