Great talk from #37c3 about prompt injections and hacking through LLMs.

https://media.ccc.de/v/37c3-12292-new_important_instructions