US national security agencies have publicly accused six Chinese AI companies of running industrial-scale campaigns to extract, or "distil", the capabilities of leading American AI models, potentially saving China billions in development costs. The NSA, CISA and FBI jointly named DeepSeek, Moonshot AI, Alibaba, MiniMax, StepFun and Z.AI, alleging they have targeted models including variants of Claude, GPT, Gemini and Grok since at least late 2024, likely with the Chinese government's awareness. The agencies say this activity threatens America's lead in the AI race and are calling for coordinated action across the industry to stop it.

The alleged tactics include mass-purchasing fraudulent accounts to run coordinated queries, sometimes numbering in the millions, and using prompt injection to trick models into revealing hidden chain-of-thought reasoning, with DeepSeek reportedly asking models to narrate their reasoning step by step. To counter this, the agencies recommend firms tighten identity verification, watch for suspicious subscription-to-usage patterns, and detect proxy-based access designed to dodge geographic restrictions. Most strikingly, they suggest US firms should covertly serve suspected malicious accounts a deliberately weaker, less capable model without notifying the user, a proposal the agencies concede would be technically difficult to implement.