Anthropic says it has cut stolen Claude reasoning tokens by up to 99% — a direct countermeasure against industrial-scale model distillation. The company confirmed the drop in a statement published Sunday, citing new detection systems and access restrictions that now block internal chain-of-thought outputs from being scraped at scale. Attackers had been using hundreds of thousands of fake accounts to extract those tokens, feeding them into rival models. Anthropic names Zhipu and Moonshot as two Chinese firms involved: Zhipu allegedly used the data to train GLM-5.3’s cyber capabilities. Moonshot is accused of routing queries through Claude without disclosure. Both models are commercially available in China and globally. Anthropic warns stolen reasoning steps could feed into foreign military or surveillance tools, not just commercial AI. So it’s now banning flagged accounts automatically. Tens of thousands have already been blocked. A classifier scans for extraction patterns in real time. And internal reasoning. The step-by-step logic behind Claude’s answers, is no longer exposed to API users or web interfaces unless explicitly enabled by enterprise customers under strict terms. That change affects developers building agents or RAG pipelines that previously relied on full reasoning traces. Legitimate users are getting caught, Anthropic admits. Its call for industry-wide adoption of similar safeguards remains voluntary. No shared standard, no enforcement mechanism. No timeline for rollout beyond ‘now live’. No pricing impact announced. No update to Claude 4, Opus 5.5, or Sonnet 4.5 models themselves. Just tighter gates around how their thinking gets seen, and who gets to see it. Anthropic says the next step is refining false-positive rates while keeping extraction volume near zero.