Generative LLM's like this often reveal confidential info that they were trained on, and anything provided to them in order to generate other output can be used as future training data which can then be re-iterated to others