Treasury Secretary Bessent told Axios that firms in China are employing "industrial distillation" to misappropriate intelligence and capabilities from artificial intelligence models created by American companies. Bessent specifically named Moonshot AI's Kimi chatbot as a powerful model that "thinks she is Claude" and claimed the system once sent People's Liberation Army weapons plans back to Anthropic. This government condemnation follows a period where Bessent observed China becoming more aware of the utility of open source models over the past 60 days.
The escalating dispute stems from a campaign OpenAI disrupted that began July 1 to extract protected reasoning from its GPT models. OpenAI attributed a core cluster of the activity to individuals associated with Moonshot AI, recording 16,000 extraction attempts from over 4,000 users during a peak on July 24 and 25. Operators manipulated model interactions to reproduce hidden reasoning without breaking encryption or breaching databases. While OpenAI closed the replay pathway on its direct services, a September 13 audit showed the vulnerability persisted for models hosted on Microsoft Azure.
Content history (2)
- 2026-10-04 · Summary · vi · Wording fixBộ trưởng Tài chính Mỹ đã gọi việc Trung Quốc thực hiện việc rút gọn mô hình AI là hành đ…Bộ trưởng Tài chính Bessent nói với Axios rằng các hãng ở Trung Quốc đang dùng "chưng cất…
- 2026-10-04 · Title · vi · Wording fixBộ trưởng Tài chính Mỹ cho rằng việc rút gọt mô hình AI của Trung Quốc là sao chép trộmBộ trưởng Tài chính Mỹ coi việc Trung Quốc chưng cất mô hình AI là ăn cắp
Key sources
- SOURCE@wallstengine“spiked to around 16,000 extraction attempts across 4,000+ users on July 24-25”x.com
- SUPPORT@firstsquawk“ACCUSING CHINESE FIRMS OF 'INDUSTRIAL DISTILLATION, WHICH IS A NICE WORD FOR STEALING FROM THE U.S. MODELS'”x.com
- SUPPORT@firstsquawk“OPERATORS DID NOT BREAK ITS ENCRYPTION, COMPROMISE A DATABASE OR GAIN DIRECT ACCESS TO STORED USER CONVERSATIONS”x.com
- SUPPORT@jschaeff3r“Patching your own API doesn’t secure your cloud-hosting ecosystem”x.com
- SUPPORT@jschaeff3r“September 13 audit found replay extraction blocked on direct OpenAI/Anthropic APIs but still working on Azure for tested OpenAI models and Opus 4.8”x.com
- SUPPORT@natolambert“manipulated model interactions so that protected reasoning could be reproduced in forms visible to the requester”x.com
- SUPPORT@pashmerepat“pewdiepie used an exploit to decrypt reasoning tokens and was banned”x.com
- SUPPORT@jessicalessin“Chinese model distillation is happening through using old work emails”x.com