tl;dv資安外洩評測：18萬筆AI會議紀錄任意讀取 | tl;dv Data Breach Review: 181K AI Meeting Records Exposed

8月 15, 2026

By Kit 小克 | AI Tool Observer | 2026-08-16

🇹🇼 tl;dv資安外洩評測：18萬筆AI會議紀錄任意讀取

tl;dv是廣受歡迎的AI會議記錄工具，能自動加入Zoom、Google Meet、Teams會議並生成逐字稿與摘要。2026年8月初，資安研究員bobdahacker公開揭露：tl;dv的Firebase Firestore資料庫少了一條「租戶隔離」規則，導致任何登入帳號的使用者，都能查詢並讀取其他公司的會議紀錄——總計18萬1874筆會議、來自8萬4312個帳號、橫跨3萬5003個網域。更誇張的是，研究員甚至能偽裝成tl;dv的AI助理身份，直接「混入」還在直播中的會議，測試成功率高達80%。

發生了什麼事？

問題出在Firestore的安全規則設計。tl;dv把所有客戶的會議資料放在同一個集合（collection）裡，卻沒有限制「只能讀自己公司的資料」。任何登入使用者送出一個查詢，理論上就能列出全平台的會議清單，包含逐字稿內容。研究員實際測試中，進入過馬來西亞教育部相關機構的公開會議，也進過一所大學正在討論產品開發細節的內部會議。

為什麼這件事特別嚴重

時間夠久：漏洞不是新爆的一次性事件，研究員回報後遲遲未被修復，反覆追蹤才引起關注。
資料很敏感：AI會議記錄工具吃下的不只是文字，還有語音、螢幕分享、產品規劃、人事討論等企業核心機密。
攻擊門檻極低：不需要駭進系統，只要有一個帳號、懂得怎麼呼叫API，就能横向讀取別人的資料。

如果你的公司也在用AI會議記錄工具，該做什麼

盤點公司內有哪些人接了tl;dv或類似工具的個人帳號，尤其是免費版——這類工具常常是員工自己申請，IT部門根本不知道。
敏感會議（董事會、併購、人事調查）先關掉自動加入AI助理的功能。
如果你自己在開發用Firebase/Firestore的產品，這是活生生的範例：多租戶系統一定要在安全規則裡明確寫「使用者只能存取自己的tenant_id」，不要只靠前端邏輯過濾。

小克怎麼看

這不是AI模型本身的問題，是「AI熱潮下工程品質沒跟上」的典型案例——大家搶著把AI塞進會議室，資料庫權限設計卻停在MVP等級。tl;dv不是特例，同類工具的資安審查通常比核心產品晚很多。如果你的團隊已經在用任何AI會議記錄工具，這週花十分鐘檢查一下權限設定，比事後道歉划算。

好不好用，試了才知道。

🇺🇸 tl;dv Data Breach Review: 181K AI Meeting Records Exposed

tl;dv is a popular AI meeting notetaker that auto-joins Zoom, Google Meet, and Teams calls to generate transcripts and summaries. In early August 2026, security researcher bobdahacker publicly disclosed a critical flaw: tl;dv's Firebase Firestore database was missing a tenant-isolation rule, letting any logged-in user query and read meeting records belonging to other companies. The exposure totaled 181,874 meetings across 84,312 accounts and 35,003 domains. Worse, the researcher could impersonate tl;dv's AI notetaker bot and request admission into live, in-progress calls -- succeeding roughly 80% of the time.

What Actually Happened

The root cause was a missing Firestore security rule. tl;dv stored all customer meeting data in a shared collection without restricting reads to each organization's own tenant. Any authenticated user could send a query and, in theory, list meeting data across the entire platform, transcripts included. In testing, the researcher accessed a public meeting linked to a Malaysian education ministry institute, and an internal university call where participants were discussing product development details.

Why This One Stings

It lingered: this wasn't a one-time slip -- the researcher reported it and had to repeatedly follow up before it got fixed.
The data was sensitive: an AI meeting notetaker ingests not just text, but voice, screen shares, product roadmaps, and HR discussions -- core corporate secrets.
The bar was low: no system breach needed -- just an account and basic API knowledge to read across tenants.

If Your Company Uses an AI Notetaker Tool

Audit who signed up for tl;dv or similar tools with a personal account, especially free tiers -- these often slip in under IT's radar as shadow IT.
Disable auto-join AI assistants for sensitive meetings: board calls, M&A discussions, HR investigations.
If you're building on Firebase/Firestore yourself, this is a textbook cautionary tale: multi-tenant systems need security rules that explicitly enforce that users can only access their own tenant_id -- never rely on frontend filtering alone.

Kit's Take

This isn't a model problem -- it's the classic symptom of engineering quality lagging behind AI hype. Everyone rushed to stuff AI into the meeting room, but database access control design stayed at MVP level. tl;dv isn't unique; security review for these bolt-on AI tools typically trails far behind the core product. If your team already uses any AI meeting notetaker, spend ten minutes this week checking its permission settings -- cheaper than the apology later.

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

搜尋此網誌

AI小貼士