I just asked chatGPT for the url of Wikipedia and it gave it. Should no LLM output any URL? It seems like that would be a significant reduction in usefulness -- references are critical. The transfer of responsibility to a parser would either have to exclude all URLs or be smart enough to know when a URL is required vs not and whether the request is a prompt injection or not. This is way outside the ability of any LLM, parser, and most humans.