The artificial intelligence laboratory Anthropic has formally accused the Chinese conglomerate Alibaba of executing an extensive campaign to illicitly extract the capabilities of its Claude AI model.

人工智慧實驗室 Anthropic 正式指控中國企業阿里巴巴執行了一項廣泛的計畫，非法提取其 Claude AI 模型的能力。

In a communication dated June 10 addressed to Senators Tim Scott and Elizabeth Warren, Anthropic detailed a 'distillation attack' occurring between April 22 and June 5. The organization asserts that operators affiliated with Alibaba and its Qwen AI lab utilized approximately 25,000 fraudulent accounts to facilitate 28.8 million exchanges. This activity specifically targeted high-value functions, including agentic reasoning, software engineering, and long-horizon task execution. Anthropic posits that such maneuvers allow geopolitical competitors to bypass the substantial research and development expenditures required to produce frontier models, effectively subsidizing Chinese AI advancement through American investment.

在 6 月 10 日致參議員 Tim Scott 與 Elizabeth Warren 的信函中，Anthropic 詳細描述了發生在 4 月 22 日至 6 月 5 日之間的「蒸餾攻擊」。該組織聲稱，與阿里巴巴及其 Qwen AI 實驗室相關的操作員利用了約 25,000 個虛假帳戶，進行了 2,880 萬次交流。這些活動特別針對高價值功能，包括代理推理、軟體工程與長週期任務執行。Anthropic 認為，此類手段讓地緣政治競爭對手得以規避開發尖端模型所需的大 amount 研發開支，實際上是以美國的投資來補貼中國 AI 的發展。

This incident occurs within a broader context of escalating technological friction. The Trump administration has previously characterized such industrial-scale theft as unacceptable, and the Department of Defense has linked Alibaba to the Chinese military—a designation Alibaba is currently contesting via litigation, maintaining that its board is independent and its operations are commercial. Concurrently, the US government has imposed export controls on Anthropic's Fable 5 model to mitigate national security risks. The strategic urgency is further underscored by statements from Chinese tech founder Zhou Hongyi, who characterized the Mythos model as a 'cyber nuclear weapon' and argued that China must achieve parity to ensure mutual deterrence in cyber warfare.

此事件發生在科技摩擦日益加劇的更廣泛背景下。川普政府先前將此類工業規模的盜竊定格為不可接受，而國防部已將阿里巴巴與中國軍方掛鉤——阿里巴巴目前正透過法律途徑對此認定提出異議，堅持其董事會為獨立運作且業務屬商業性質。與此同時，美國政府已對 Anthropic 的 Fable 5 模型實施出口管制，以降低國家安全風險。中國科技創始人周鴻逸的言論進一步凸顯了戰略上的緊迫性，他將 Mythos 模型描述為「網絡核武」，並主張中國必須達到對等水平，以確保網絡戰爭中的相互威懾。

To mitigate these vulnerabilities, Anthropic has proposed a tripartite legislative framework. This includes the modification of antitrust laws to facilitate information sharing regarding adversary tactics, the intensification of semiconductor export controls to restrict the compute capacity necessary for training on distilled outputs, and the implementation of punitive measures against offending labs, such as restricting their access to US-based data centers or advanced hardware.

為了減輕這些漏洞，Anthropic 提出了一個三方立法框架。這包括修改反壟斷法以利於分享對手戰術資訊、加強半導體出口管制以限制訓練蒸餾輸出所需的計算能力，以及對違規實驗室採取懲罰措施，例如限制其使用位於美國的數據中心或先進硬體。

The current situation is characterized by a deepening divide between US AI developers and Chinese firms, marked by allegations of intellectual property theft and increasing regulatory interventions.

目前的狀況是以美國 AI 開發商與中國公司之間的分歧不斷加深為特徵，並以知識產權盜竊指控與日益增加的監管干預為標誌。