ChatGPT 讓 Cloudflare 偷讀你的資料：打字前已收集 55 項資訊 | ChatGPT Lets Cloudflare Collect 55 Data Points Before You Type
By Kit 小克 | AI Tool Observer | 2026-03-30
🇹🇼 ChatGPT 讓 Cloudflare 偷讀你的資料：打字前已收集 55 項資訊
ChatGPT 的機器人防護系統，比你想的侵入性更強。一位開發者近日解密了 Cloudflare 部署在 ChatGPT 上的混淆 JavaScript，發現在你輸入第一個字之前，系統已靜默收集了 55 項關於你的資料——從 GPU 型號、IP 地理位置，到你的 React 應用程式狀態，全部一網打盡。這篇調查報告在 HackerNews 登上榜首，獲得 650 分以上及 413 則熱烈討論。
Cloudflare Turnstile 在 ChatGPT 上做什麼？
使用 Cloudflare 做機器人防護本身並不罕見。問題在於這次調查揭露的深度：Cloudflare 的 Turnstile 程式不只檢查你的 IP 或 User-Agent，它直接讀取 ChatGPT 的 React 應用程式內部狀態。
程式碼中，它存取三個 React 內部屬性：__reactRouterContext、loaderData、clientBootstrap。這是為了確認「真正的 React 程式碼確實在瀏覽器中執行完畢」，而非只是載入靜態 HTML。換句話說，這是史上少見的應用程式層機器人偵測，而非傳統的網路或瀏覽器層操作。
55 項資料：三層收集機制
解密後的腳本揭示，Turnstile 共收集 55 個屬性，分三個層次：
- 瀏覽器層：GPU 型號與效能、螢幕解析度、硬體規格、已安裝字型列表、本地儲存空間
- 網路層：城市名稱、IP 地址、緯度經度、連接 IP、使用者地區（由 Cloudflare 伺服器端注入）
- 行為層（透過 Signal Orchestrator）：鍵盤輸入時機間隔、滑鼠移動速度與軌跡、捲動模式、貼上事件
更重要的是：這份「指紋」（fingerprint）會以編碼金鑰存入瀏覽器的 localStorage，跨頁面持久追蹤你，而非每次重新運算。
為什麼這比普通的防機器人措施更嚴重？
傳統 Cloudflare 防護只需在網路或瀏覽器層運作，不需要碰 React 狀態。這次的設計選擇讓第三方服務（Cloudflare）能讀取 OpenAI 應用程式的內部資料結構——這從架構上就是一個重大的隱私邊界突破。
社群的另一個批評更犀利：ChatGPT 的訓練資料本身就大量來自未經授權的網路爬取，如今卻用如此侵入性的手段防止他人爬取自己的服務——這個諷刺沒有逃過 HackerNews 的眼睛。
OpenAI 的隱私政策中，幾乎沒有針對 Cloudflare Turnstile 資料收集範圍的清晰說明。
你能做什麼？
- 使用 uBlock Origin 嘗試攔截 Cloudflare 腳本（但可能導致 ChatGPT 功能異常）
- 改用 OpenAI API 直接呼叫，避開瀏覽器介面的追蹤
- 考慮本地端開源模型（Ollama + Llama 4、Mistral 等）以完全迴避這個問題
- 定期清除瀏覽器 localStorage，移除已存的追蹤指紋
AI 工具越來越強大，但它們背後的資料收集機制也越來越複雜。身為用戶，我們有必要了解自己在付出什麼代價。好不好用，試了才知道。
🇺🇸 ChatGPT Lets Cloudflare Collect 55 Data Points Before You Type
The bot-protection system running on ChatGPT is more invasive than you probably knew. A developer recently decrypted the obfuscated JavaScript that Cloudflare deploys on ChatGPT and discovered it silently collects 55 data properties before you type your first character — including GPU specs, your precise geographic location, and ChatGPT's internal React application state. The technical writeup hit #1 on Hacker News with 650+ points and 413 comments.
What Is Cloudflare Turnstile Actually Reading?
Using Cloudflare for bot protection is standard. What's unusual here is the depth. Cloudflare's Turnstile script doesn't just check your IP or User-Agent — it directly reads ChatGPT's internal React application state.
Specifically, it accesses three React internals: __reactRouterContext, loaderData, and clientBootstrap. The purpose is to confirm that actual React application code executed in your browser — not just that HTML loaded. This makes it a rare case of application-layer bot detection, which goes far deeper than traditional network or browser-layer checks.
55 Data Points: Three Collection Layers
The decrypted script collects 55 properties across three layers:
- Browser layer: GPU model and capabilities, screen resolution, hardware specs, installed fonts, local storage details
- Network layer: City name, IP address, latitude/longitude, connecting IP, user region (injected server-side by Cloudflare)
- Behavioral layer (via Signal Orchestrator): Keystroke timing intervals, mouse movement speed and trajectory, scroll patterns, paste events
This fingerprint is then stored in your browser's localStorage under an encoded key, persisting across page loads to enable continuous tracking — not recalculated each session.
Why This Crosses a Line
Traditional Cloudflare protection operates at the network or browser layer — it has no need to read React application state. This design choice allows a third-party service (Cloudflare) to read OpenAI's internal application data structures. Architecturally, that's a meaningful privacy boundary violation.
The HackerNews community also raised a sharper critique: ChatGPT's training data was itself built largely on unauthorized web scraping, yet OpenAI now uses aggressively invasive techniques to prevent others from scraping their own service. The irony wasn't lost.
Notably, OpenAI's privacy policy contains almost no disclosure about the scope of data that Cloudflare Turnstile collects on their behalf.
What Users and Developers Can Do
- Use uBlock Origin to attempt blocking Cloudflare scripts (may break ChatGPT functionality)
- Access OpenAI via the API directly to avoid browser-based tracking entirely
- Consider locally-hosted open-source models (Ollama + Llama 4, Mistral, etc.) to sidestep this entirely
- Periodically clear your browser's localStorage to remove stored fingerprints
AI tools are growing more capable — but so is the surveillance infrastructure embedded within them. As users and developers, it's worth understanding exactly what we're trading away when we use "free" AI services.
好不好用，試了才知道。
Sources / 資料來源
- ChatGPT won't let you type until Cloudflare reads your React state (原始技術分析)
- Hacker News 討論串 (650+ 分，413 則留言)
延伸閱讀 / Related Articles
- Claude Code Bug：背景自動執行 git reset --hard 靜默刪除你的程式碼 | Claude Code Bug: Background git reset --hard Silently Deletes Your Code
- ARC-AGI-3 發布：頂尖 AI 全部得分不到 1% | ARC-AGI-3: Every Top AI Model Scored Under 1%
- OpenAI Spud 完成訓練：GPT-6 幾週後上線？ | OpenAI Spud Training Done: Is GPT-6 Weeks Away?
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