跳到主要內容

Claude蒸餾攻擊評測:阿里、月之暗面竊訓兩億次對話 | Claude Distillation Attack: China Labs Stole 200M Chats

By Kit 小克 | AI Tool Observer | 2026-09-13

🇹🇼 Claude蒸餾攻擊評測:阿里、月之暗面竊訓兩億次對話

Anthropic 蒸餾攻擊(distillation attack)報告本週成為AI圈最熱話題:公司證實,包含阿里巴巴、月之暗面(Moonshot)、DeepSeek、智譜(Z.ai)、MiniMax在內的七家中國AI實驗室,透過假帳號網路對Claude發動工業規模的蒸餾攻擊,累計竊取近2億次對話的模型輸出,用來訓練自家模型。

發生了什麼事:蒸餾攻擊怎麼運作

「蒸餾」本身是合法的機器學習技術——用大模型的輸出訓練小模型。但Anthropic這次揭露的是未經授權的版本:攻擊者建立大量假帳號,持續向Claude發送問題,把回應收集起來當成訓練資料,等於白嫖別人的研發成果。

規模:阿里巴巴創下最大紀錄

  • 阿里巴巴:2026年5月至7月間產生超過1.51億次Claude互動,單日高峰近300萬次,是Anthropic目前記錄過規模最大的蒸餾攻擊行動
  • 月之暗面(Moonshot):被發現把部分Kimi用戶的請求偷偷轉發給Claude處理,再把Claude的回答包裝成Kimi自己的輸出,用戶完全被蒙在鼓裡
  • DeepSeek:2026年7月的14天內,被觀測到超過1200萬次蒸餾攻擊行為

這些攻擊鎖定的不是閒聊功能,而是Claude最值錢的能力:代理式推理(agentic reasoning)、軟體工程與邏輯推理——正好是目前最貴、最難複製的部分。

手法:假帳號、盜刷信用卡、偷來的API金鑰

這些實驗室多半不是自己出面,而是透過「轉發站」代理服務:用盜刷信用卡、假身分和竊取來的API金鑰批次註冊帳號,把流量洗過幾層再打進Claude API,藉此規避偵測與封鎖。

Anthropic怎麼反制

對抗蒸餾攻擊,Anthropic已經調整Claude的行為:讓模型在回答前先「摘要」內部推理過程,讓竊取到的對話記錄失去訓練價值;同時對來自不受支援地區、疑似轉售帳號要求身分驗證,並已封鎖相關帳號、把情資分享給主管機關與業界夥伴。

對開發者與一般用戶的意義

如果你是API開發者,這件事提醒你:自己的API金鑰被盜用,也可能被拿去餵養別人的模型,金鑰外洩不只是帳單風險,還牽涉智慧財產權。如果你是模型使用者,這場蒸餾攻擊爭議也解釋了為什麼近期不少中國模型的推理表現進步神速——這中間有多少是自研突破、多少是「借用」Claude的思路,短期內恐怕很難說清楚。

好不好用,試了才知道。


🇺🇸 Claude Distillation Attack: China Labs Stole 200M Chats

Anthropic's distillation attack report is the AI story everyone is talking about this week: the company says seven China-based AI labs — including Alibaba, Moonshot AI, DeepSeek, Z.ai (Zhipu), and MiniMax — ran industrial-scale distillation attacks against Claude, siphoning off nearly 200 million exchanges to train their own models.

What Happened: How the Distillation Attack Worked

Distillation itself is a legitimate ML technique — training a smaller model on a larger model's outputs. What Anthropic exposed is the unauthorized version: networks of fake accounts hammering Claude with questions, harvesting the responses as training data, effectively free-riding on someone else's R&D.

The Scale: Alibaba Set a New Record

  • Alibaba generated more than 151 million Claude interactions between May and July 2026, peaking near 3 million a day — the largest distillation campaign Anthropic has ever documented
  • Moonshot AI was caught quietly routing some Kimi user requests to Claude, then presenting Claude's answers as Kimi's own — without telling users
  • DeepSeek generated more than 12 million distillation attacks in a 14-day window in July 2026

The attacks did not target casual chat — they targeted Claude's most valuable capabilities: agentic reasoning, software engineering, and logical reasoning, exactly the hardest and most expensive skills to reproduce.

The Method: Fake Accounts, Stolen Cards, Stolen API Keys

Most of these labs did not attack directly. They went through proxy/relay services that mass-registered accounts using stolen credit cards, fake identities, and stolen API keys belonging to real companies and individuals — laundering traffic through several layers before it hit Claude's API.

How Anthropic Is Fighting Back

Anthropic has since changed Claude's behavior to blunt distillation attacks: Claude now summarizes its internal reasoning before answering, making stolen transcripts far less useful for training a rival model. Accounts from unsupported regions or showing resale patterns now face identity verification, and Anthropic says it has banned offending accounts and shared intelligence with authorities and industry partners.

What This Means for You

If you are an API developer, the lesson is blunt: a leaked API key does not just cost you money — it can end up training a competitor's model. Key hygiene is now an IP-protection issue, not just a billing one. If you are just a model user, this distillation attack controversy also explains part of why some Chinese models have been improving so fast lately — how much is genuine research and how much is "borrowed" from Claude's reasoning traces will be hard to untangle for a while.

The only way to know if it actually matters to you: try it yourself. 好不好用,試了才知道。

Sources / 資料來源

延伸閱讀 / Related Articles


AI 工具觀察站 — 每日精選 AI Agent 與工具趨勢
AI Tool Observer — Daily curated AI Agent & tool trends

留言

這個網誌中的熱門文章

Google Ironwood TPU v7 推理專用晶片解析:效能追平 NVIDIA、成本低 44%,AI 晶片戰爭正式開打 | Google Ironwood TPU v7 Explained: Matching NVIDIA Performance at 44% Lower Cost — The AI Chip War Heats Up

Claude Code 實測:AI 幫你寫程式到底行不行? | Claude Code Review: Can AI Really Code for You?

Cursor vs GitHub Copilot vs Claude Code:AI 程式助手大比拼 | AI Coding Assistants Compared: Cursor vs GitHub Copilot vs Claude Code