The NSA, FBI, and CISA have issued a joint advisory accusing six Chinese AI firms of systematically stealing intellectual property from American frontier models through industrial-scale data distillation.

Key Points

• The advisory identifies DeepSeek, Moonshot AI, Alibaba, MiniMax, StepFun, and Z.AI as companies conducting unauthorized data extraction campaigns.
• Targeted U.S. models include GPT, Claude, Gemini, and Grok, with firms allegedly using fraudulent accounts and API proxies to bypass security restrictions.
• Distillation techniques involve training cheaper models on the outputs of more capable systems, including hidden chain-of-thought reasoning and reinforcement learning data.
• The report claims DeepSeek’s low training costs were achieved by excluding the value of data acquired through these malicious distillation efforts.
• U.S. agencies recommend that AI providers deploy subtle output degradation for suspected malicious users to protect proprietary model capabilities.

Why it Matters

This technical advisory marks a significant escalation in U.S. efforts to curb the unauthorized transfer of AI technology to Chinese firms. By providing specific evidence of model exploitation, the government is signaling that future trade sanctions or stricter export controls may be imminent.