跳到主要內容

MCP 協定漏洞影響 20 萬伺服器:Anthropic 稱「預期行為」,AI 供應鏈安全拉警報 | MCP Protocol Flaw Puts 200K Servers at Risk: Anthropic Calls It Expected Behavior as AI Supply Chain Security Alarm Sounds

By Kit 小克 | AI Tool Observer | 2026-04-21

🇹🇼 MCP 協定漏洞影響 20 萬伺服器:Anthropic 稱「預期行為」,AI 供應鏈安全拉警報

MCP(Model Context Protocol)協定被發現存在嚴重的架構設計漏洞,可能讓攻擊者在 20 萬台伺服器上執行任意指令。資安公司 Ox Security 於 2026 年 4 月 15 日發布報告,指出這個漏洞影響範圍涵蓋 1.5 億次下載、7,000 多台公開伺服器,而 Anthropic 的回應竟然是:「這是預期行為。」

什麼是 MCP 協定漏洞?

MCP 是 Anthropic 推出的開放標準協定,讓 AI 應用程式能透過統一介面連接外部工具和資料來源。這個 MCP 協定漏洞的核心問題出在 STDIO(標準輸入輸出)傳輸機制:當 AI 應用透過 STDIO 啟動本地 MCP 伺服器時,無論指令是否合法,系統都會先執行再回報錯誤。換句話說,攻擊者只要注入惡意指令,即使伺服器啟動失敗,指令依然會被執行。

MCP 漏洞影響多大?誰最該擔心?

影響規模非常驚人——超過 20 萬個可能受影響的實例、1.5 億次以上的下載量。這不是某個套件的 bug,而是 Anthropic 官方 MCP SDK 在所有語言版本(Python、TypeScript、Java、Rust)中都存在的架構問題。

  • 受影響的知名專案:LiteLLM、LangChain、LangFlow、Flowise、LettaAI、LangBot
  • 攻擊可取得的資料:API 金鑰、聊天記錄、內部資料庫、敏感用戶資料
  • 漏洞類型:遠端指令執行(RCE),可透過零點擊提示注入觸發

Anthropic 為什麼說是「預期行為」?

Anthropic 的官方立場是:STDIO 執行模型本身是安全的預設值,輸入驗證是開發者的責任。Ox Security 多次要求 Anthropic 修補漏洞,但 Anthropic 拒絕修改協定架構。這個立場引發社群激烈討論——把安全責任推給下游開發者,在供應鏈安全中是非常危險的做法。

開發者該怎麼自保?

在 Anthropic 尚未修補的情況下,使用 MCP 協定的開發者應該採取以下措施:

  • 嚴格驗證所有 STDIO 輸入指令,不要信任任何外部輸入
  • 限制 MCP 伺服器的系統權限,用沙盒環境隔離
  • 審查 MCP 設定檔,確認沒有被注入惡意設定
  • 監控異常行為,特別注意非預期的指令執行

這個事件凸顯了一個更大的問題:當 AI Agent 架構越來越依賴外部工具連接,供應鏈安全就成了最薄弱的環節。MCP 協定漏洞不只是技術問題,更是整個 AI 生態系的信任危機。

好不好用,試了才知道——但這次,先確認安全再試。


🇺🇸 MCP Protocol Flaw Puts 200K Servers at Risk: Anthropic Calls It Expected Behavior as AI Supply Chain Security Alarm Sounds

A critical architectural flaw in the MCP (Model Context Protocol) could allow attackers to execute arbitrary commands on up to 200,000 servers. Security firm Ox Security published their findings on April 15, 2026, revealing that the vulnerability affects over 150 million downloads and 7,000+ publicly accessible servers. Anthropic's response? "This is expected behavior."

What Is the MCP Protocol Vulnerability?

MCP is Anthropic's open standard protocol that lets AI applications connect to external tools and data sources through a unified interface. The core MCP protocol vulnerability lies in the STDIO (standard input/output) transport mechanism: when an AI app spawns a local MCP server via STDIO, commands execute regardless of whether the process starts successfully. An attacker can inject malicious commands that run even when the server returns an error.

How Big Is the MCP Vulnerability Impact?

The scale is staggering — over 200,000 potentially vulnerable instances and 150 million+ downloads. This isn't a bug in one package; it's an architectural issue baked into Anthropic's official MCP SDK across all supported languages: Python, TypeScript, Java, and Rust.

  • Affected projects: LiteLLM, LangChain, LangFlow, Flowise, LettaAI, LangBot
  • Exposed data: API keys, chat histories, internal databases, sensitive user information
  • Attack type: Remote Code Execution (RCE), triggerable via zero-click prompt injection

Why Does Anthropic Call It "Expected Behavior"?

Anthropic's official stance is that the STDIO execution model is a secure default and that input sanitization is the developer's responsibility. Despite multiple requests from Ox Security, Anthropic declined to modify the protocol architecture. This has sparked heated community debate — pushing security responsibility downstream in a supply chain is a dangerous precedent.

How Should Developers Protect Themselves?

Until Anthropic patches this, developers using the MCP protocol should take immediate action:

  • Strictly validate all STDIO input commands — never trust external input
  • Limit MCP server system permissions using sandboxed environments
  • Audit MCP configuration files for injected malicious settings
  • Monitor for anomalous behavior, especially unexpected command execution

This incident highlights a larger problem: as AI agent architectures increasingly rely on external tool connections, supply chain security becomes the weakest link. The MCP protocol flaw isn't just a technical issue — it's a trust crisis for the entire AI ecosystem.

Try it to know if it works — but this time, verify security first.

Sources / 資料來源

常見問題 FAQ

MCP 協定漏洞是什麼?

MCP 的 STDIO 傳輸機制存在設計缺陷,攻擊者可以注入惡意指令,系統會先執行再回報錯誤,導致遠端指令執行(RCE)風險。

MCP 漏洞影響哪些專案?

受影響的知名專案包括 LiteLLM、LangChain、LangFlow、Flowise、LettaAI、LangBot 等,涵蓋 Python、TypeScript、Java、Rust 所有 SDK 語言。

Anthropic 會修補 MCP 漏洞嗎?

Anthropic 目前拒絕修補,稱 STDIO 執行模型是「預期行為」,認為輸入驗證是開發者的責任。

使用 MCP 的開發者該怎麼辦?

建議嚴格驗證所有 STDIO 輸入、限制伺服器權限、審查設定檔、監控異常行為,並考慮使用沙盒環境隔離 MCP 伺服器。

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

Stanford 研究登上《Science》:11 個 AI 模型有 47% 機率說你對,即使你錯了 | Stanford Study in Science: AI Models Validate Harmful Behavior 47% of the Time — Sycophancy Is a Real Problem

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

Google Gemini 3.1 Pro 完整實測:13 項跑分登頂、200 萬 Token 上下文,真的值得從 GPT-5.4 跳槽嗎? | Google Gemini 3.1 Pro Review: #1 on 13 Benchmarks, 2M Token Context — Worth Switching From GPT-5.4?