Anthropic, a San Francisco-based AI company, has accused three prominent Chinese AI laboratories – DeepSeek, Moonshot AI, and MiniMax – of orchestrating large-scale campaigns to extract capabilities from its Claude models using tens of thousands of fraudulent accounts. These alleged campaigns represent concrete evidence of foreign competitors using distillation, a process of knowledge extraction from powerful AI models, to accelerate their own research and development.
Distillation, while a legitimate training method, can be weaponized to capture capabilities developed by others. Anthropic’s technical blog post detailed how these Chinese labs generated millions of exchanges with Claude, targeting specific capabilities like agentic reasoning and coding. The use of proxy networks and ‘hydra cluster’ architectures allowed the labs to bypass access restrictions set by Anthropic, posing significant national security risks.
Anthropic’s response includes building detection systems, sharing indicators with industry players, and calling for coordinated action. The company’s revelations are expected to impact ongoing policy debates, including chip export controls and API security considerations across the AI industry. The era of treating model access as a simple transaction may be evolving into a landscape where API security is paramount.
Source: VentureBeat