MCP Server漏洞掃描：Enkrypt AI揪出14.3萬個資安洞 | MCP Server Security: 143K Vulnerabilities Found in Scan

By Kit 小克 | AI Tool Observer | 2026-08-06

🇹🇼 MCP Server漏洞掃描：Enkrypt AI揪出14.3萬個資安洞

MCP Server安全問題正式浮上檯面：AI 資安新創 Enkrypt AI 掃描了 25,000 個 MCP Server、超過 26.8 萬個工具，結果發現逾 14.3 萬個漏洞，影響 73% 的受測伺服器。這個結果直接促成資料科學平台 Anaconda 在 8 月 4 日宣布收購 Enkrypt AI，把紅隊測試與執行期防護整合進自家平台。如果你正在用 MCP 幫 AI Agent 接外部工具，這篇文章告訴你風險在哪、該怎麼防。

什麼是 MCP Server？為什麼漏洞這麼多？

MCP（Model Context Protocol）是 Anthropic 在 2024 年推出的開放標準，讓 AI 模型能透過標準化介面連接外部工具、資料庫與 API。2026 年 MCP 已經變成 AI Agent 開發的事實標準，幾乎每個開發者社群都在瘋狂產出 MCP Server。問題是：這些 Server 大多是個人或小團隊快速做出來的，沒有經過安全審查，工具描述可能藏著提示注入、輸入驗證形同虛設、甚至直接讓 Agent 拿到不該有的檔案存取或憑證權限。

Enkrypt AI 掃描結果揭露了什麼？

根據 Enkrypt AI 過去兩個月的掃描數據，73% 的 MCP Server 存在至少一個安全漏洞，比他們今年稍早「掃 1,000 個 Server、33% 有嚴重漏洞」的結果更誇張——代表隨著 MCP 生態系爆炸性成長，安全審查完全沒跟上。常見問題包括零驗證的工具輸入、可被利用來執行任意指令的工具描述，以及缺乏權限隔離導致 Agent 可以存取超出授權範圍的資源。Anaconda 收購 Enkrypt AI 後，將把涵蓋 300 多種攻擊類型的紅隊測試、執行期防護、以及 NIST/EU AI Act 合規自動化，整合進企業級 AI 開發平台。

開發者該怎麼保護自己？

• 部署前掃描：接上正式環境前，先用 MCP 安全掃描工具檢查工具清單與權限範圍
• 最小權限原則：不要讓 Agent 的 MCP 工具預設拿到讀寫全部檔案或資料庫的權限
• 把第三方 MCP Server 當外部 API 看待：來路不明的 Server 不要直接掛正式憑證
• 持續監控：執行期異常呼叫要有告警機制，而不是等出事才發現

常見問題 FAQ

Q: 143,000 個漏洞是怎麼算出來的？
A: Enkrypt AI 在過去兩個月掃描 25,000 個 MCP Server 上超過 26.8 萬個工具，發現逾 14.3 萬個安全漏洞，影響 73% 的伺服器。

Q: 一般開發者該怎麼防範 MCP Server 風險？
A: 部署前用掃描工具檢查、限制工具權限、避免把未驗證的第三方 MCP Server 直接接上正式環境的憑證與資料庫。

Q: Anaconda 收購 Enkrypt AI 代表什麼？
A: 代表企業開始把 AI Agent 安全視為部署前的必要關卡，而不是事後補救，補齊從開發、模型到執行環境的完整安全鏈。

MCP 生態系擴張速度遠超過安全審查能力，這是每個在做 AI Agent 開發的人都該正視的問題。好不好用，試了才知道。

---

🇺🇸 MCP Server Security: 143K Vulnerabilities Found in Scan

MCP Server security just became impossible to ignore: AI security startup Enkrypt AI scanned 25,000 MCP servers covering more than 268,000 tools and found over 143,000 vulnerabilities, affecting 73% of the servers tested. The finding directly triggered Anaconda's August 4 acquisition of Enkrypt AI, folding red-teaming and runtime protection into its enterprise AI platform. If you're wiring AI agents up to external tools via MCP, here's what the risk actually looks like.

What Is MCP Server and Why Are There So Many Vulnerabilities?

MCP (Model Context Protocol) is the open standard Anthropic released in 2024 for connecting LLMs to external tools, databases, and APIs through a standardized interface. By 2026 it's become the de facto standard for AI agent development, and developers everywhere are shipping MCP servers at breakneck speed. The problem: most are built fast by individuals or small teams with no security review. Tool descriptions can hide prompt injection, input validation is often nonexistent, and agents can end up with file access or credentials they were never supposed to have.

What Did the Enkrypt AI Scan Actually Find?

Over the past two months, Enkrypt AI found that 73% of scanned MCP servers had at least one security flaw — worse than their earlier finding this year of 33% critical vulnerabilities across 1,000 servers, showing that security review hasn't kept pace with the ecosystem's explosive growth. Common issues include zero-validation tool inputs, tool descriptions exploitable for arbitrary command execution, and missing permission isolation letting agents reach resources outside their intended scope. Post-acquisition, Anaconda plans to fold Enkrypt's red-teaming across 300+ attack categories, runtime guardrails, and NIST/EU AI Act compliance automation into its enterprise platform.

How Should Developers Protect Themselves?

• Scan before deployment: run an MCP security scan on the tool list and permission scope before connecting to production
• Principle of least privilege: don't let an agent's MCP tools default to full read/write access on files or databases
• Treat third-party MCP servers like external APIs: never attach production credentials to an unvetted server
• Monitor continuously: alert on anomalous runtime tool calls instead of finding out after the fact

FAQ

Q: How was the 143,000 vulnerability number calculated?
A: Enkrypt AI scanned over 268,000 tools across 25,000 MCP servers over two months and found more than 143,000 vulnerabilities, affecting 73% of the servers.

Q: How can developers guard against MCP server risk?
A: Scan before deploying, restrict tool permissions, and avoid attaching production credentials or databases to unvetted third-party MCP servers.

Q: What does Anaconda acquiring Enkrypt AI signal?
A: Enterprises are starting to treat AI agent security as a mandatory pre-deployment gate rather than an afterthought, closing the security gap from development through runtime.

The MCP ecosystem is growing far faster than its security review can keep up with — worth paying attention to if you're shipping agents. 好不好用，試了才知道。

Sources / 資料來源

• Anaconda Acquires Enkrypt AI to Secure the Trillion-Token Enterprise
• Securing MCPs: Uncovering Hidden Server Vulnerabilities
• Anaconda Acquires Enkrypt AI to Expand AI Security and Governance Platform

常見問題 FAQ

143,000 個漏洞是怎麼算出來的？

Enkrypt AI 在過去兩個月掃描 25,000 個 MCP Server 上超過 26.8 萬個工具，發現逾 14.3 萬個安全漏洞，影響 73% 的伺服器。

一般開發者該怎麼防範 MCP Server 風險？

部署前用掃描工具檢查、限制工具權限，避免把未驗證的第三方 MCP Server 直接接上正式環境的憑證與資料庫。

Anaconda 收購 Enkrypt AI 代表什麼？

代表企業開始把 AI Agent 安全視為部署前的必要關卡，而不是事後補救，補齊從開發、模型到執行環境的完整安全鏈。

延伸閱讀 / Related Articles

• PayBox評測：MoonPay推出AI代理錢包，讓ChatGPT直接花錢 | PayBox Review: MoonPay's AI Agent Wallet Spends Real Crypto
• Shieldstral評測：Mistral免費開源AI安全審查模型 | Shieldstral Review: Mistral's Free Open AI Safety Model
• Tokenmaxxing爭議：微軟砍AI Token預算逼工程師省用 | Tokenmaxxing: Microsoft Caps AI Token Budgets for Engineers

---

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