I think that almost all use cases for LLMs that process untrusted inputs are unsafe. See https://greshake.github.io/ and https://github.com/greshake/llm-security for more information.