Claude drops research that just 250 malicious documents (roughly 420k tokens, representing 0.00016% of total training tokens can cook a model! Let the manipulation begin! https://www.anthropic.com/research/small-samples-poison