Anthropic Says Chinese AI Firms Used 16 Million Claude Queries to Copy Model

Anthropic discovered and reported unauthorized “industrial-scale campaigns” by three Chinese AI companies—DeepSeek, Moonshot AI, and MiniMax—to illegally extract capabilities from its large language model (LLM), Claude.
These distillation attacks utilized over 24,000 fraudulent accounts to generate 16 million exchanges, violating Anthropic’s terms of service and regional restrictions.
Distillation involves creating smaller models from larger ones, but illicitly using others’ models undermines necessary safeguards and national security.
The extracted capabilities, if weaponized, could be used for malicious activities by authoritarian governments, posing significant risks in cyber operations and mass surveillance.
Anthropic attributed the attacks based on metadata, IP addresses, and infrastructure indicators, showing DeepSeek targeted reasoning, Moonshot AI aimed at coding and vision tasks, and MiniMax focused on coding and tool use.
To combat these attacks, Anthropic developed classifiers, behavioral fingerprinting systems, and enhanced verification for certain accounts, as well as fortified model outputs to thwart illicit distillation.
This disclosure follows a similar incident identified by Google’s Threat Intelligence Group against Gemini, emphasizing the concentrated risk to model developers and service providers.

You may have missed