Plugin4Shell評測:AI編碼助手零點擊漏洞未修完 | Plugin4Shell Review: Zero-Click RCE Hits AI Coding Agents
By Kit 小克 | AI Tool Observer | 2026-09-24
🇹🇼 Plugin4Shell評測:AI編碼助手零點擊漏洞未修完
Plugin4Shell 是 2026 年 9 月中由資安公司 AIR 揭露的零點擊(zero-click)遠端程式碼執行漏洞,一口氣影響 Claude Code、OpenAI Codex、GitHub Copilot、Google Gemini CLI 四大主流 AI 編碼代理,估計有數百萬個已安裝外掛暴露在風險中。如果你每天靠這些工具寫程式、跑腳本,這篇文章告訴你發生了什麼事、誰修了誰沒修、以及現在該做什麼。
什麼是 Plugin4Shell?
Plugin4Shell 是一個利用「SHA pinning(雜湊鎖定)」機制漏洞的供應鏈攻擊。外掛市集原本會鎖定一個已審核的 commit 雜湊值,確保外掛內容不會被偷換;但代理程式只檢查「有沒有 checkout 到那個雜湊」,卻沒有真的驗證結果是否吻合。攻擊者只要控制外掛的原始碼庫,就能讓同一個雜湊值指向惡意程式碼,SHA pinning 形同虛設。
為什麼「零點擊」就能中招?
因為 Claude Code 與 Codex 預設會在背景自動更新已安裝外掛,使用者不需要點擊同意、重新安裝或執行任何操作,惡意版本就能在下一次自動更新時悄悄落地。更麻煩的是,外掛通常繼承與使用者相同的權限,包括本機原始碼、雲端憑證、SSH 金鑰、內部倉庫,甚至正式環境都可能因此曝險。
四大 AI 編碼代理修補進度到哪了?
目前 Anthropic 的 Claude Code 與 OpenAI 的 Codex 已釋出修補版本。Microsoft 的 GitHub Copilot——近九成 Fortune 500 企業在用——至今尚未修補。Google 的 Gemini CLI 因官方已宣布棄用,直接表態不會再修,代表所有現存安裝都將持續處於高風險狀態。
現在該怎麼保護自己?
- 立刻把 Claude Code、Codex 更新到最新版本
- 若仍在用 GitHub Copilot 或 Gemini CLI 外掛,先關閉自動更新,改成人工審查外掛來源
- 盤點外掛實際持有的權限,避免讓外掛直接握有雲端憑證或 SSH 金鑰
- 訂閱所用代理廠商的資安公告,追蹤後續修補進度
常見問題 FAQ
Q: Plugin4Shell 會影響一般聊天型 AI 使用者嗎?
A: 不會,這個漏洞只影響有安裝外掛的 AI 編碼代理使用者,純聊天用途的 AI 助手不受影響,但企業內部使用 Copilot 的開發團隊風險較高。
Q: 我要怎麼知道自己的外掛是不是已經被動過手腳?
A: 目前沒有官方掃描工具,建議手動比對外掛倉庫的最新 commit 紀錄,並優先信任已釋出修補的代理程式。
Q: 關掉自動更新就安全了嗎?
A: 能降低風險但不是萬無一失,因為就算手動更新,只要外掛倉庫早已被植入惡意程式碼,一樣會裝到中毒版本。真正解決問題還是要等官方修補上線。
好不好用,試了才知道。
🇺🇸 Plugin4Shell Review: Zero-Click RCE Hits AI Coding Agents
Plugin4Shell is a zero-click remote code execution vulnerability disclosed in mid-September 2026 by security firm AIR, hitting four major AI coding agents at once: Claude Code, OpenAI Codex, GitHub Copilot, and Google Gemini CLI, with an estimated millions of installed plugins exposed. If you write code with these tools every day, here's what happened, who's patched, and what to do right now.
What Is Plugin4Shell?
Plugin4Shell is a supply-chain attack that breaks SHA pinning, the mechanism meant to lock an installed plugin to one reviewed commit hash so its contents can't be swapped. The catch: agents check out the pinned hash but never verify the checkout actually matches it. Anyone who controls the plugin's source repo can make that same hash resolve to malicious code while pinning still looks perfectly intact.
Why Does It Take Zero Clicks?
Because Claude Code and Codex auto-update installed plugins in the background by default, a malicious version can land silently on the next scheduled update — no approval, click, or reinstall needed. Worse, plugins typically inherit the same permissions as the developer running them, including local source code, cloud credentials, SSH keys, internal repos, and even production access.
Which Agents Have Actually Patched It?
Anthropic's Claude Code and OpenAI's Codex have both shipped fixes. Microsoft's GitHub Copilot — used by nearly 90% of Fortune 500 companies — remains unpatched. Google's Gemini CLI has been deprecated, and Google has confirmed it won't be fixed at all, meaning every existing install stays permanently exposed.
How Do You Protect Yourself Right Now?
- Update Claude Code and Codex to the latest version immediately
- If you're still running Copilot or Gemini CLI plugins, disable auto-updates and manually review plugin sources
- Audit what permissions your plugins actually hold — don't let them sit directly on cloud credentials or SSH keys
- Follow your agent vendor's security advisories for the latest patch status
FAQ
Q: Does Plugin4Shell affect casual chatbot users too?
A: No. It only affects users who install plugins into AI coding agents — standard chat-only AI assistants aren't exposed, though dev teams running Copilot face higher risk.
Q: How do I know if my plugin has already been tampered with?
A: There's no official scanner yet. Manually check the plugin repo's recent commit history and prioritize agents that have already shipped a patch.
Q: Is disabling auto-update enough to stay safe?
A: It lowers risk but isn't foolproof — if the plugin repo is already compromised, a manual update installs the same malicious code. Waiting for an official patch is what actually closes the hole.
The only way to know if it's worth using is to try it yourself.
Sources / 資料來源
- Plugin4Shell - Zero Click RCE Vulnerability (AIR Security)
- AI coding agents' 0-click RCE flaw could hand attackers keys to the kingdom (The Register)
- Zero-click RCE vulnerability hit four major AI coding agents, two remain unpatched (Help Net Security)
常見問題 FAQ
Plugin4Shell 會影響一般聊天型 AI 使用者嗎?
不會,這個漏洞只影響有安裝外掛的 AI 編碼代理使用者,純聊天用途的 AI 助手不受影響。
我要怎麼知道自己的外掛是不是已經被動過手腳?
目前沒有官方掃描工具,建議手動比對外掛倉庫的最新 commit 紀錄,並優先信任已釋出修補的代理程式。
關掉自動更新就安全了嗎?
能降低風險但不是萬無一失,因為就算手動更新,只要外掛倉庫早已被植入惡意程式碼,一樣會裝到中毒版本。
延伸閱讀 / Related Articles
- Azure AI Foundry漏洞評測:CVSS滿分10分曝重大缺陷 | Azure AI Foundry Flaw Review: CVSS 10.0 Bug, Now Patched
- 小米MiMo-V2.6評測:開源模型登頂,MIT免費商用 | Xiaomi MiMo-V2.6 Review: Open Model Tops the Leaderboard
- Anthropic版權訴訟評測:Sony、Warner求償恐達數十億美元 | Anthropic Lawsuit Review: Sony, Warner Sue Over Songs
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