跳到主要內容

Big Sleep AI評測:Google揪出Chrome藏13年漏洞 | Big Sleep AI Review: Google Finds 13-Year Chrome Bug

By Kit 小克 | AI Tool Observer | 2026-08-07

🇹🇼 Big Sleep AI評測:Google揪出Chrome藏13年漏洞

Google 最近公布一個很難忽視的數字:由 DeepMind 與 Project Zero 合作打造的 AI 資安代理 Big Sleep,在 Chrome 最近兩個穩定版 release 裡,協助揪出並修補了 1,072 個資安漏洞,其中一個嚴重的沙盒逃逸漏洞(sandbox escape)已經藏在程式碼裡超過 13 年都沒被發現。這不是實驗室展示的漂亮數字,而是真的被塞進正式版的修補量。

什麼是 Big Sleep AI?

Big Sleep AI 是一套用大型語言模型驅動的自動化資安代理,任務是在 Chromium 這種龐大到人力難以窮盡的程式碼庫裡,主動搜尋潛在漏洞、重現問題、評估嚴重程度,甚至直接生成候選修補程式。

Big Sleep 怎麼運作?

整套流程分三段:先讓模型理解程式碼邏輯、推測可能出錯的邊界條件;接著實際嘗試觸發崩潰或異常行為,驗證漏洞是否真實存在;最後交給人類資安團隊複審,決定要不要採用 AI 生成的修補建議。Google 強調 AI 目前仍是輔助角色,關鍵漏洞的最終判斷還是人類把關。

Big Sleep 修復了哪些關鍵漏洞?

  • Chrome 最近兩個穩定版 release 中,Big Sleep AI 協助發現並修補 1,072 個資安漏洞
  • 其中一個嚴重的沙盒逃逸漏洞已藏在程式碼裡 13 年未被發現
  • 2026 年 5 月單月,Big Sleep 攔下超過 20 個漏洞流入正式版,包含一個嚴重等級問題
  • Google 另有 Gemini 驅動的代理框架,專門在更廣的 Chromium 程式碼庫掃描漏洞、降低誤判率

對開發者和一般用戶有什麼影響?

對一般用戶來說,Chrome 更新頻率不會變,但每次更新堵住的洞可能比過去多,等於瀏覽器的隱形防護在悄悄變強。對開發者跟資安團隊而言,這是個明確訊號:AI 輔助程式碼審查已經不是概念驗證,而是能量產真實修補程式的正式流程,大型專案的漏洞回應速度可能會被重新定義。

小克實測心得

老實說,1,072 這個數字聽起來很像行銷話術,但真正值得注意的是那個藏了 13 年的沙盒逃逸漏洞——這代表傳統靠人力審查、fuzzing 的方式真的有盲區。Big Sleep AI 補上的是「量」,不是取代人類的「判斷」。如果你是資安或維運團隊,現在該問的不是要不要導入 AI 輔助掃描,而是怎麼把 AI 的輸出納入既有複審流程,千萬別把 AI 生成的修補直接 merge 進 production。

好不好用,試了才知道。


🇺🇸 Big Sleep AI Review: Google Finds 13-Year Chrome Bug

Google just published a number that's hard to ignore: its AI security agent Big Sleep, built with DeepMind and Project Zero, helped find and fix 1,072 security bugs across Chrome's last two stable releases — including a sandbox escape vulnerability that had gone undetected in the codebase for 13 years. This isn't a lab demo figure; these are real patches that shipped to production.

What Is Big Sleep AI?

Big Sleep is an LLM-driven security agent that autonomously hunts for vulnerabilities across the massive Chromium codebase, reproduces bugs, triages severity, and even drafts candidate patches — compressing work that used to take human researchers weeks into an automated pipeline.

How Does Big Sleep Actually Work?

The pipeline runs in three stages: the model reasons about code logic to spot likely edge cases, then actively tries to trigger crashes or abnormal behavior to confirm a real vulnerability exists, and finally hands everything to human reviewers who decide whether to accept the AI-generated fix. Google is clear that AI remains an assistant — humans still make the final call on critical bugs.

What Vulnerabilities Did Big Sleep Actually Fix?

  • Across Chrome's two most recent stable releases, Big Sleep AI helped find and patch 1,072 security bugs
  • One critical sandbox escape had gone undetected in the codebase for 13 years
  • In May 2026 alone, Big Sleep blocked over 20 vulnerabilities from reaching production, including one critical-severity issue
  • Google also runs a separate Gemini-powered agent harness that scans the broader Chromium codebase with fewer false positives

What Does This Mean for Developers and Everyday Users?

For everyday users, Chrome's update cadence won't change, but each release is likely plugging more holes than before — invisible protection quietly getting stronger. For developers and security teams, the signal is clear: AI-assisted code review for vulnerability hunting has moved past proof-of-concept and into a workflow that ships real patches at scale.

Kit's Honest Take

1,072 sounds like a marketing number, but what actually matters is the 13-year-old sandbox escape — proof that traditional human review and fuzzing genuinely have blind spots. Big Sleep AI adds volume, not judgment; it extends human review, it doesn't replace it. If you run a security or platform team, the real question isn't whether to adopt AI-assisted scanning — it's how to fold its output into your existing review process without auto-merging AI-generated patches straight to production.

好不好用,試了才知道 — Kit

Sources / 資料來源

常見問題 FAQ

Big Sleep AI 是什麼?

Big Sleep 是 Google 與 DeepMind、Project Zero 合作開發的 AI 資安代理,用大型語言模型自動在 Chromium 程式碼庫中尋找、驗證並協助修補資安漏洞。

Big Sleep 找到了多少個漏洞?

在 Chrome 最近兩個穩定版 release 中,Big Sleep 協助發現並修補了 1,072 個資安漏洞,其中一個沙盒逃逸漏洞已藏在程式碼裡 13 年未被發現。

AI 修補的程式碼會直接上線嗎?

不會。Google 強調 AI 生成的修補建議仍須經過人類資安團隊複審才會採用,AI 目前扮演輔助角色,不是取代人類判斷。

一般 Chrome 使用者需要做什麼?

不需要額外操作,只要保持 Chrome 自動更新,就能持續獲得這些 AI 協助發現的漏洞修補。

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