Anthropic says it cut Claude reasoning token theft by up to 99% in fight against industrial-scale distillation

Distillation is when a rival records a model's answers at scale and trains a cheaper copy on them. Anthropic says defenses that hide Claude's internal reasoning steps and flag odd query patterns cut that harvesting by 90 to 99 percent, banning tens of thousands of accounts. Anthropic links the attacks to Chinese firms Zhipu and Moonshot.

Summary written with AI by HowTechMoves, based on reporting by The Decoder.