“Anthropic's report on distillation attacks by Chinese companies noted that one of the methods used was employing Anthropic's models as an "LLM as a judge" for other model outputs.”