Anthropic strengthens defences against distillation attacks against Claude, cuts data harvesting by upto 99%

Anthropic's new defenses primarily target the internal reasoning transcripts of its Claude models, which are particularly valuable for distillers. These transcripts represent the AI's thinking process before generating a final response, making them crucial for replicating advanced capabilities in competitor models. The company describes distillation as covert campaigns that use fraudulent accounts to mimic frontier AI for a fraction of the typical development time and cost.

The report highlights that unauthorized distillation poses significant biosecurity, cybersecurity, and national security risks, as distilled models often lack adequate safety guardrails. Anthropic cited examples, including Chinese laboratory Zhipu's GLM-5.3, which allegedly extracted cyber capabilities from leading American models, and Moonshot's Kimi models routing queries to Claude, including one linked to China's People's Liberation Army.

Anthropic's defensive framework operates on three fronts: controlling reasoning data movement, deploying a specialized two-stage classifier with a reported false-positive rate below 0.1 percent, and restricting persistent abusers through account bans, usage caps, and billing for blocked queries. Implementing these safeguards carries commercial and technical trade-offs, with Anthropic acknowledging it has turned away significant revenue and trillions of tokens in traffic from malicious networks.

The focus on protecting reasoning data rather than just final outputs suggests a deeper understanding of how sophisticated AI models are reverse-engineered. This approach acknowledges that the underlying thought process, not merely the end product, is the true intellectual property being targeted. The cited examples underscore the strategic implications of such intellectual property theft, extending beyond commercial competition to national security concerns.