What happened?
Anthropic announced it has cut the leakage of Claude's internal reasoning tokens by 90 to 99 percent. The company links these large-scale "distillation" attacks to Chinese AI firms Zhipu and Moonshot.
What is distillation?
Distillation means collecting a powerful model's outputs and using them to train a smaller model. According to Anthropic, some companies are doing this at industrial scale: they try to capture not just Claude's final answers but its internal reasoning steps.
What Anthropic did
The company says it detected and blocked such access and banned tens of thousands of accounts. It reports a 90-99 percent drop in token theft after these measures.
Why it matters
Internal reasoning data from large language models can make it easier for rivals to build similar capabilities more cheaply. Anthropic's move signals that model providers will take stricter measures to protect their own outputs.
