I remember doing some work on this on GPT-2. Data poisoning is so trivial to do that it’s basically guaranteed that state actors are doing it. They just have to

I remember doing some work on this on GPT-2. Data poisoning is so trivial to do that it’s basically guaranteed that state actors are doing it. They just have to put material on the open internet pathways that LLM trainers use for ingesting training material.