Azure AI Foundry漏洞評測：CVSS滿分10分曝重大缺陷 | Azure AI Foundry Flaw Review: CVSS 10.0 Bug, Now Patched

9月 23, 2026

By Kit 小克 | AI Tool Observer | 2026-09-24

🇹🇼 Azure AI Foundry漏洞評測：CVSS滿分10分曝重大缺陷

Azure AI Foundry 這幾天在資安圈炸了鍋——微軟自家的 AI 代理開發平台被抓到一個 CVSS 滿分 10 分的漏洞（CVE-2026-85889），攻擊者不需要任何帳密就能直接對內部管理介面下指令，一路提權拿下控制權。這是目前少數幾個拿到滿分評分的雲端 AI 平台漏洞，也讓「AI 代理平台的資安門檻夠不夠高」重新被拿出來討論。

CVE-2026-85889 是什麼：漏洞成因拆解

根據微軟官方安全公告，這個漏洞被歸類為 CWE-306（關鍵功能缺少身份驗證）。簡單說，Azure AI Foundry（前身為 Azure AI Studio）某些內部 RESTful API 端點，完全沒有檢查 session token、API key 或 OAuth2 bearer 憑證，攻擊者只要組出對的 HTTP 請求，就能直接呼叫管理層級的功能，等於繞過整套權限驗證機制。

影響範圍：用來建構、部署、管理生成式 AI 應用與 Agent 的企業平台
發現者：資安研究員 Rémy Marot（@R_Marot）
是否已被利用：目前沒有證據顯示曾被實際攻擊利用
修補狀態：微軟已在伺服器端全面修復，使用者不需要額外操作

為什麼這個 AI 代理平台漏洞比一般 CVE 更值得注意

一般雲端服務的漏洞頂多影響單一功能，但 Azure AI Foundry 是拿來管理 AI Agent 部署與權限的中樞平台。一旦被攻破，攻擊者理論上能碰到的不只是資料，還包括企業內部串接的模型、知識庫、甚至下游自動化流程。這也是為什麼多家資安媒體（The Hacker News、Cloud Security Alliance）都特別點出：AI 代理基礎設施的攻擊面正在快速擴大，而且往往比傳統雲端服務更難稽核，因為權限鏈路又長又複雜。

企業該做什麼

雖然微軟已伺服器端修復，仍建議檢查 Azure AI Foundry 的存取日誌，確認過去是否有異常呼叫紀錄
重新盤點內部 Agent 平台的權限設計，不要假設「平台廠商做好驗證」就等於安全
建立獨立於平台之外的稽核機制，尤其是能觸發部署或提權操作的 API

這次事件沒有造成實際災情，算是有驚無險，但它提醒所有正在導入 AI Agent 的團隊：平台的資安基本功（身份驗證）一旦漏掉，AI 帶來的自動化能力反而會放大災情範圍。好不好用，試了才知道。

🇺🇸 Azure AI Foundry Flaw Review: CVSS 10.0 Bug, Now Patched

Azure AI Foundry, Microsoft's platform for building and deploying AI agents, just had a maximum-severity security scare: a CVSS 10.0 vulnerability (CVE-2026-85889) that let unauthenticated attackers reach internal management APIs and escalate privileges — no credentials required. It is one of the few AI platform bugs ever to hit a perfect severity score, and it has reopened the question of whether AI agent infrastructure is being secured as carefully as it is being shipped.

What CVE-2026-85889 Actually Is

Microsoft classifies the flaw as CWE-306, missing authentication for a critical function. In plain terms, certain internal RESTful endpoints in Azure AI Foundry never validated session tokens, API keys, or OAuth2 bearer credentials. An attacker who crafted the right HTTP request could call management-level functions directly, bypassing the platform entire authorization layer.

Affected service: Azure AI Foundry (formerly Azure AI Studio), used to build, deploy, and manage generative AI apps and agents
Discovered by: security researcher Rémy Marot (@R_Marot)
Exploited in the wild? No evidence found so far
Fix status: patched server-side by Microsoft — no customer action required

Why This AI Agent Platform Bug Matters More Than a Typical Cloud CVE

Most cloud vulnerabilities compromise a single feature. Azure AI Foundry sits at the center of how enterprises deploy and govern AI agents — models, permissions, connected knowledge bases, and downstream automations all route through it. A missing-authentication bug there does not just leak data; it potentially hands over the controls for everything the platform orchestrates. That is why outlets like The Hacker News and the Cloud Security Alliance flagged it specifically as a sign that AI agent infrastructure attack surface is growing faster than the tooling to audit it.

What Teams Running AI Agents Should Actually Do

Even with the server-side fix live, check your Azure AI Foundry access logs for anomalous calls during the exposure window
Re-audit permission design for internal agent platforms — assuming the vendor handles auth is not a security control
Build audit logging independent of the platform itself, especially for any API that can trigger deployment or privilege changes

No real-world damage was reported this time, so call it a near miss. But it is a reminder for anyone rolling out AI agents in production: skip authentication basics, and the automation AI agents bring just means a bigger blast radius when something breaks. 好不好用，試了才知道 — you only really know once you have tried it yourself.

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

搜尋此網誌

AI小貼士