跳到主要內容

MLflow漏洞評測:CVSS 9.8,預設設定就被拿下root | MLflow RCE Review: CVSS 9.8 Bug Gives Root by Default

By Kit 小克 | AI Tool Observer | 2026-10-29

🇹🇼 MLflow漏洞評測:CVSS 9.8,預設設定就被拿下root

如果你的團隊用MLflow追蹤機器學習或LLM實驗,現在該停下來檢查版本了。本週曝光的CVE-2026-116692是一個CVSS評分9.8的高危漏洞,攻擊者不需要登入就能在預設設定的MLflow伺服器上遠端執行任意程式碼,而且官方Docker映像檔是以root身分在跑,等於一次拿下整台機器。

漏洞原理:MLflow怎麼被「反序列化」玩壞

問題出在MLflow 3.12版新增的「第三方Scorer」功能。原本的驗證邏輯只檢查模組路徑是否「等於」或「以某個允許的前綴開頭」,而不是鎖死成一組固定白名單。攻擊者可以利用MLflow自己的artifact上傳功能,把一個惡意Python模組放進被允許的路徑底下,再誘使伺服器把它當成Scorer匯入並執行——整個過程完全不需要驗證身分。

受影響版本

  • MLflow 3.12.0rc0 到 3.16.1 都受影響
  • 官方v3.14.0 Docker映像檔預設以root執行,中獎等於整台伺服器淪陷
  • 目前已有公開可用的PoC(概念驗證)程式碼在流傳

誰該緊張?MLOps與LLMOps團隊的現實風險

MLflow不是小眾工具,它是目前最常見的機器學習與LLM實驗追蹤平台之一,很多團隊圖方便直接用預設設定跑在內網甚至公開雲端,沒開身分驗證。這代表只要伺服器能被外部連到,攻擊者就能靠這個漏洞直接拿到伺服器執行權限,進一步接觸訓練資料、模型權重、API金鑰等敏感內容。

修補方法

好消息是修法很單純:

  • 升級到MLflow 3.17.0或以上版本:pip install --upgrade "mlflow>=3.17.0"
  • 如果暫時不能升級,至少先幫MLflow伺服器加上身分驗證、關掉對外網路存取
  • 檢查是否已經有不明的Scorer或artifact被上傳過,留意是否已遭入侵

這類漏洞再次證明一個老生常談:AI工具鏈的安全債,往往比模型本身的能力債更容易被忽略。MLflow這麼基礎的基礎設施出現CVSS 9.8等級的漏洞,不是因為AI多危險,而是因為大家把它當成內部玩具,沒有用正式服務的標準去防護。好不好用,試了才知道。


🇺🇸 MLflow RCE Review: CVSS 9.8 Bug Gives Root by Default

MLflow just got hit with a nasty one: CVE-2026-116692, a CVSS 9.8 remote code execution bug that lets an unauthenticated attacker run arbitrary Python code on a default MLflow server — and on the official Docker image, that code runs as root. If your team tracks ML or LLM experiments with MLflow, this is worth stopping to check right now.

How the Bug Works: Scorer Deserialization Gone Wrong

The flaw lives in MLflow 3.12s third-party scorer feature. The validation logic checked whether a module path equals or starts with an allowed prefix, instead of matching against a closed, exact whitelist. An attacker can abuse MLflows own artifact upload endpoint to drop a malicious Python module under an allowed path, then trick the server into importing and instantiating it as a scorer — no login required.

Who is Affected

  • MLflow versions 3.12.0rc0 through 3.16.1
  • The official v3.14.0 Docker image runs as root by default, so a successful exploit means full server takeover
  • Public proof-of-concept exploit code is already circulating

Why MLOps and LLMOps Teams Should Care

MLflow is one of the most widely used experiment-tracking platforms for both classic ML and LLM workflows, and plenty of teams run it with default settings — no authentication, sometimes reachable from outside the internal network. That is a direct path for an attacker to reach training data, model weights, and API keys sitting on the same box.

The Fix

  • Upgrade to MLflow 3.17.0 or later: pip install --upgrade "mlflow>=3.17.0"
  • If you cannot upgrade immediately, put authentication in front of the server and block public network access
  • Audit for unexpected scorers or artifacts uploaded during the exposure window

The lesson here is not that AI is scary — it is that core AI infrastructure like MLflow often gets deployed with the security posture of an internal toy, not a production service. A CVSS 9.8 bug in something this foundational is a reminder to treat your MLOps stack like any other internet-facing system. You won't know until you try it — 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