Training a model on another model's answers is called distillationTraining a small, cheap model to imitate a big one's outputs, keeping much of the capability at a fraction of the cost.Full definition →, and it has legitimate uses. Anthropic's February 2026 report alleges a different use: coordinated access through fraudulent accounts to collect Claude outputs for competing models. It attributes campaigns to DeepSeek, Moonshot AI and MiniMax, citing request metadata, infrastructure patterns and account coordination.
The detection problem is collective behavior. An individual request can resemble normal use, while repeated requests across linked accounts concentrate on material useful for training. Anthropic's account supports an allegation of organized extraction; it does not independently establish how much any resulting model improved.
Terms in this piece · Glossary
distillation — Training a small, cheap model to imitate a big one's outputs, keeping much of the capability at a fraction of the cost.
