Nice experiment and write-up on how to backdoor LLM. Have a read. https://open.substack.com/pub/shrivu/p/how-to-backdoor-large-language-models