The NSA, CISA, and FBI say several China-based AI companies have been extracting knowledge from U.S. frontier models through large-scale distillation, affecting models from Claude, GPT, Gemini, and Grok. The agencies warn this activity threatens U.S. technological leadership, with DeepSeek and Moonshot AI among the companies named in the report. #DeepSeek #MoonshotAI #Alibaba #MiniMax #StepFun #ZAI #Claude #GPT #Gemini #Grok
Keypoints
- China-based AI firms are accused of distilling knowledge from U.S. frontier models at scale.
- DeepSeek reportedly used Claude, GPT-4, GPT-5, Gemini, and Grok 4 data to train R1 and V3.
- Moonshot AI is said to have used Claude Fable 5 and GPT-4o outputs to improve Kimi models.
- The agencies mapped the tactics to MITRE ATLAS and highlighted novel evasion and optimization techniques.
- The report says the campaign poses a strategic threat to U.S. economic interests and technological leadership.