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

Anthropic disclosed that it has reduced the theft of Claude's internal reasoning tokens by between 90 and 99 percent. The company identified large-scale distillation campaigns as the primary threat, in which attackers systematically extract the model's chain-of-thought outputs to train competing systems.

The company linked these attacks to Chinese AI firms Zhipu and Moonshot, and said it has already banned tens of thousands of accounts involved in the activity. This represents one of the first public acknowledgments by a major AI lab of organized, industrial-scale model distillation targeting its reasoning capabilities.

Distillation attacks are a growing concern in the AI industry because they allow competitors to replicate proprietary model behavior without building from scratch. By extracting reasoning tokens, attackers can train smaller models that mimic the target's internal logic, potentially eroding the intellectual property advantage of frontier labs.

The disclosure highlights a broader tension in the AI ecosystem: as frontier models become more capable, the incentive to extract and replicate their capabilities through API abuse grows. Anthropic's response—combining technical defenses with account enforcement—may set a precedent for how other labs address similar threats going forward.

This page provides an editorial summary based on publicly available information. It is not a republished article. Use the source link below for the original report.