GitLab AI Gateway漏洞評測:CVSS 9.9,樣板跳脫變RCE | GitLab AI Gateway RCE Review: CVSS 9.9 Sandbox Escape
By Kit 小克 | AI Tool Observer | 2026-10-19
🇹🇼 GitLab AI Gateway漏洞評測:CVSS 9.9,樣板跳脫變RCE
GitLab AI Gateway在2026年10月2日被揭露一個CVSS評分高達9.9的重大漏洞CVE-2026-90970,攻擊者只要擁有Duo Agent Platform的帳號權限,就能透過精心設計的流程設定,跳脫Prompt樣板沙箱,在AI Gateway主機上執行任意指令。這是今年至今評分最高的AI工具鏈漏洞之一,也再次提醒:把LLM塞進CI/CD流程裡,攻擊面不會變小,只會變得更隱蔽。
漏洞成因:樣板引擎裡藏著指令注入
這顆漏洞被歸類為CWE-1336(樣板引擎特殊字元處理不當)。GitLab AI Gateway在組裝送給模型的Prompt時,會把Duo Agent Platform「自訂流程(custom flow)」設定值代入樣板。問題在於,系統沒有把這些使用者可控的輸入徹底隔離——一名已登入且擁有Duo Agent Platform權限的使用者,只要送出特別構造的流程設定,就能突破原本應該隔離Prompt組裝邏輯的沙箱,直接在Gateway服務的作業系統層執行指令。說白了,這就是經典的伺服端樣板注入(SSTI),只是這次穿著「AI Agent」的外衣。
受影響範圍:自架使用者才要緊張
- 受影響版本:18.1.6–19.2.3、19.3.0–19.3.1、19.4.0
- GitLab.com、GitLab Dedicated,以及接到GitLab官方代管Gateway的Self-Managed使用者:已自動修補,不用做任何事
- 自行架設AI Gateway(Self-Hosted AI Gateway)的企業:必須盡快升級,否則Duo Agent Platform等於開了一個遠端程式碼執行的後門
修補版本
GitLab已釋出AI Gateway 19.2.4、19.3.2、19.4.1三個修補版本,自架環境應立即升級到對應版本線。
小克怎麼看
這起事件和近期一連串AI工具鏈資安事故同一個套路:功能先上線,權限模型後補。GitLab AI Gateway的CVSS 9.9分不是因為攻擊技巧多高深——樣板注入是教科書等級的老問題——而是因為它插在CI/CD的核心位置,一旦被打穿,影響範圍是整條交付鏈。如果公司有自架AI Gateway,先確認版本號,再盤點誰有Duo Agent Platform的存取權,別等到被掃到才知道。
🇺🇸 GitLab AI Gateway RCE Review: CVSS 9.9 Sandbox Escape
On October 2, 2026, GitLab disclosed CVE-2026-90970, a critical CVSS 9.9 vulnerability in its AI Gateway that let an authenticated Duo Agent Platform user escape the prompt template sandbox and execute arbitrary commands on the gateway host. It is one of the highest-severity AI tooling bugs disclosed so far this year, and another reminder that wiring LLMs into your CI/CD pipeline does not shrink your attack surface — it just makes it harder to see.
Root Cause: Injection Hiding Inside a Template Engine
The flaw is tracked as CWE-1336 — improper neutralization of special elements in a template engine. GitLab AI Gateway assembles prompts sent to the model by substituting values from Duo Agent Platform "custom flow" configurations into templates. The problem: those user-controlled values were not fully isolated. An authenticated user with Duo Agent Platform access could submit a crafted flow configuration that broke out of the sandbox meant to contain prompt assembly, resulting in command execution at the OS level on the AI Gateway service itself. Strip away the AI framing and it is a textbook server-side template injection (SSTI) — just dressed up as an agent feature.
Who Actually Needs to Worry
- Affected versions: 18.1.6–19.2.3, 19.3.0–19.3.1, 19.4.0
- GitLab.com, GitLab Dedicated, and Self-Managed instances pointed at GitLab's hosted AI Gateway: already patched automatically, no action needed
- Organizations running their own Self-Hosted AI Gateway: must upgrade now — otherwise Duo Agent Platform access is effectively a remote code execution backdoor
Patched Versions
GitLab shipped fixes in AI Gateway 19.2.4, 19.3.2, and 19.4.1. Self-hosted operators should upgrade to the matching patch line immediately.
Kit's Take
This follows the same pattern as the recent string of AI toolchain security incidents: ship the feature first, backfill the permission model later. The CVSS 9.9 here is not about exotic exploitation — template injection is an old, well-understood bug class — it is about placement. This component sits inside the CI/CD core, so a successful exploit touches the entire delivery chain. If your org self-hosts an AI Gateway, check the version number first, then audit who actually holds Duo Agent Platform access. Do not wait for a scanner to find it for you.
好不好用,試了才知道。
Sources / 資料來源
- GitLab AI Gateway Sandbox Escape Leads to Full RCE
- CVE-2026-90970: Critical GitLab AI Gateway Flaw Fixed
- GitLab CVE-2026-90970: Critical AI Gateway RCE Flaw
延伸閱讀 / Related Articles
- DeepSeek搜尋API評測:千次0.75美元,比Tavily便宜10倍 | DeepSeek Search API Review: $0.75/1K vs Tavily's $8
- AI代理人評測:六大銀行封殺,密碼授權誰負責 | AI Agent Review: Banks Block Rogue Agents Over Liability
- OpenAI人形機器人評測:2027年demo,身體不是重點 | OpenAI Humanoid Robot Review: 2027 Demo, Brain Over Body
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
張貼留言