Data poisoning is a critical issue, especially in regulated domains. I find the recent research by Afforai on techniques for testing and mitigating data poisoning in LLMs to be quite insightful.