A small number of samples can poison LLMs of any size Anthropic

Anthropic research on data-poisoning attacks in large language models