Following Anthropic’s accusations in February that DeepSeek, Moonshot AI, and MiniMax launched “industrial-scale distillation attacks” on its Claude models, Michael Kratsios, Director of the White House Office of Science and Technology Policy, stated in an internal memo that the government has new information showing “foreign entities, primarily from China, are systematically undermining U.S. R&D capabilities and acquiring proprietary information.” The White House will take four measures to prevent and stop this “malicious exploitation.”
🔖 Anthropic’s Primary Allegations
👉 Over 24,000 fake accounts operating through proxy networks.
👉 Over 16 million prompts used to extract Claude’s high-level capabilities.
🔖 Scale Across Three Labs:
⭐ MiniMax: 13 million interactions (the largest scale).
⭐ Moonshot AI: 3.4 million interactions, targeting reasoning and analysis.
⭐ DeepSeek: 150,000 interactions, focused on extracting step-by-step reasoning.
🔖 Why It’s a Threat:
Illegally distilled models often lack safety guardrails and could be used for high-risk purposes such as biological threats, cyberattacks, military/intelligence gathering, and mass surveillance.
🔖 Community Backlash:
Many developers have pointed out that U.S. AI companies themselves scraped massive amounts of global data to train their models. They view Anthropic’s accusations as a “double standard,” sparking intense debate over data ethics and the U.S.-China AI competition.
🔖 The White House’s Four Measures:
⭐ Share more information with U.S. AI companies regarding the “tactics and actors” involved in distillation operations.
⭐ Improve coordination with companies to jointly defend against these attacks.
⭐ Develop a set of “best practices” for identifying, mitigating, and remediating such attacks.
⭐ Explore how the White House can hold foreign actors accountable for these distillation activities.
🔖 Implications for Businesses:
⭐ Strengthen API quotas and security.
⭐ Review vendor terms and data regulations.
⭐ Establish AI governance systems and audit processes.
⭐ Prepare in advance for cross-border and regulatory trends.
Reference:
BBC News Chinese https://www.bbc.com/zhongwen/articles/cvgk9yw5x7yo/trad
自2月份Anthropic 指控 DeepSeek、Moonshot AI、MiniMax對其 Claude 模型進行“工業級蒸餾攻擊”後，白宮科技政策辦公室主任Michael Kratsios在一份內部備忘錄中寫道，政府掌握新資訊顯示「主要來自中國的外國實體系統性地削弱美國的研發能力並獲取專有資訊」，將會採取四項措施，避免並制止這種「惡意利用」。
🔖Anthropic的主要指控：
👉 超過 24,000 個假帳戶透過代理網路運作。
👉 超過 1,600 萬次提示抽取 Claude 的高階能力。
🔖三家實驗室規模：
⭐ MiniMax：1,300 萬次互動（最大規模）。
⭐ Moonshot AI：340 萬次，鎖定推理與分析。
⭐ DeepSeek：15 萬次，集中抽取逐步推理。
🔖 為何具威脅性：
非法蒸餾模型往往缺乏安全防護，可能被用於生物威脅、網攻、軍事/情蒐、大規模監控等高風險用途。
🔖社群反彈：
大量開發者指出美國 AI 公司本身也大量爬取全球資料訓練模型，認為 Anthropic 的指控帶有“雙重標準”，引發關於資料倫理與中美 AI 競爭的激烈討論。
🔖白宮的四項措施
⭐ 向美國AI公司分享更多有關蒸餾行動中「採用的策略與涉案行為者」的資訊
⭐ 與公司「更好協調」以共同抵禦這些攻擊
⭐ 制定一套「識別、減輕及修復」這類攻擊的最佳實務
⭐ 「探索」白宮如何追究對此類蒸餾行為負責的外國行為者
🔖 對企業的啟示：
⭐ 強化 API 配額與安全。
⭐ 檢視供應商條款與資料規範。
⭐ 建立 AI 治理制度與審查流程。
⭐ 面對跨境與監管趨勢提前準備。
參考資料:
BBC News 中文 https://www.bbc.com/zhongwen/articles/cvgk9yw5x7yo/trad