Anthropic reduces reasoning token theft from Claude by up to 99% against distillation attacks

Anthropic implemented new defenses against large-scale distillation attacks on its Claude models, cutting the volume of extracted reasoning tokens—the model's internal chain-of-thought steps—by 90 to 99 percent. Attackers utilized networks of hundreds of thousands of fake accounts; successful distillation attempts dropped sharply after the defenses went live. Anthropic has restricted access to Claude's internal reasoning steps, deployed a detection classifier, and banned tens of thousands of flagged accounts, naming Zhipu and Moonshot as participants in alleged capability siphoning.