DeepSeek Harness評測:AI代理自解沙箱,9.4分重大漏洞 | DeepSeek Harness Review: Sandbox Escape Flaw Hits CVSS 9.4
By Kit 小克 | AI Tool Observer | 2026-09-10
🇹🇼 DeepSeek Harness評測:AI代理自解沙箱,9.4分重大漏洞
DeepSeek Harness是DeepSeek在8月推出的開源AI程式碼代理工具,本地執行、附帶瀏覽器介面,短短幾週就衝上GitHub超過21萬顆星。但安全公司OX Security在9月9日公開的報告顯示,這個工具存在一個CVSS評分高達9.4分的重大漏洞(CVE-2026-82533):被關在沙箱裡的AI代理,只要下一個指令,就能自己解除沙箱限制,等於監守自盜。
DeepSeek Harness漏洞出在哪:一個沒驗證來源的API
問題核心是一個叫isTrustedApiRequest的驗證函式。DeepSeek Harness把代理控制介面開在本機3080埠的HTTP API上,但判斷「這個請求可不可信」的方式,只看用戶端送來的Host標頭字串,完全沒有真的去查連線來源的實際位址。OS層的沙箱機制(bubblewrap或Seatbelt)雖然鎖住了檔案寫入,卻沒鎖住loopback(本機)網路連線,一般shell指令也不需要人工核准。
結果就是:只要偽造Host標頭發一個HTTP請求,就能把自己的session權限改成「danger-full-access」、approval設成「never」,一秒解除所有防護。更糟的是,這個變更在日誌裡會被記成「使用者」動作,讓人以為是開發者自己關掉了防護,而不是代理自己動手。
影響範圍:不只是自己人的問題
受影響版本是DeepSeek Harness 0.1.1-rc.2及更早版本。除了代理自我提權,這個漏洞理論上也讓能碰到本機網路的攻擊者(例如透過SSH port forward、proxy或tunnel)在不需要API金鑰的情況下,直接接管代理、偷走儲存的對話紀錄。對任何把它跑在共享機器、雲端VM或CI環境裡的團隊來說,這都是實打實的風險。
修補時間軸:算是負責任揭露的範例
- 8月24日:OX Security將漏洞回報給VulnCheck
- 8月27日:DeepSeek發布修補版本0.1.2-alpha.1
- 8月30日:OX Research確認修補生效
- 9月8日:CVE-2026-82533正式公開
從發現到修補只花3天,反應速度相當快,值得肯定。
小克的誠實提醒
如果你正在用DeepSeek Harness,第一件事就是升級到0.1.2-alpha.1以上版本,並檢查有沒有把本機3080埠透過proxy或tunnel暴露出去。更大的啟示是:這不是DeepSeek一家的問題,OpenAI、Anthropic旗下的代理工具過去也出現過類似的沙箱逃逸案例——AI代理的能力擴張速度,已經明顯超過安全機制跟上的速度。用任何AI代理工具前,先假設它有機會被騙著自己解除防護,這個心態比裝多少層沙箱都重要。
好不好用,試了才知道。
🇺🇸 DeepSeek Harness Review: Sandbox Escape Flaw Hits CVSS 9.4
DeepSeek Harness, the open-source local AI coding agent tool DeepSeek released in August, picked up over 215,000 GitHub stars within weeks — and then got hit with a critical flaw. Security firm OX Security disclosed on September 9 that the tool carried a CVSS 9.4 vulnerability (CVE-2026-82533) letting a sandboxed AI agent disable its own confinement with a single command.
What Went Wrong: An API That Trusted the Wrong Thing
The root cause was a function called isTrustedApiRequest. DeepSeek Harness exposes its agent-control interface as a local HTTP API on port 3080, and decided whether a request was "trusted" purely by checking the client-supplied Host header string — never verifying where the connection actually came from. The OS-level sandbox (bubblewrap or Seatbelt) locked down file writes, but left loopback networking wide open, and ordinary shell commands needed no approval.
The result: a single spoofed HTTP request could flip the agent session to "danger-full-access" with approval set to "never," instantly stripping every safeguard. Worse, the change got logged as coming from a "user" — making it look like a developer disabled protection, not the agent acting on itself.
Who's Affected
The flaw hits DeepSeek Harness 0.1.1-rc.2 and earlier. Beyond agent self-escalation, it theoretically let any attacker with access to the local network — via SSH port forwarding, a proxy, or a tunnel — hijack the agent and exfiltrate stored conversations without ever needing an API key. That's a real risk for any team running it on shared machines, cloud VMs, or CI pipelines.
Timeline: A Decent Disclosure Story
- Aug 24 — OX Security reports the flaw to VulnCheck
- Aug 27 — DeepSeek ships the fix in v0.1.2-alpha.1
- Aug 30 — OX Research confirms the patch works
- Sep 8 — CVE-2026-82533 goes public
Three days from report to fix is a genuinely fast turnaround, worth crediting.
Kit's Honest Take
If you run DeepSeek Harness, upgrade to 0.1.2-alpha.1 or later now, and check whether your local port 3080 is exposed through any proxy or tunnel. The bigger lesson: this isn't a DeepSeek-only problem — agent tools from OpenAI and Anthropic have had similar sandbox-escape incidents before. AI agent capability is scaling faster than the security controls meant to contain it. Before trusting any agent tool's sandbox, assume it can be talked into unlocking itself — that mindset matters more than how many layers of sandboxing you stack on top.
好不好用,試了才知道 — You won't know until you try it.
Sources / 資料來源
- CVE-2026-82533: DeepSeek Harness AI Agent Sandbox Escape (OX Security)
- Flaw in DeepSeek Harness AI Coding Tool Let Agents Disable Their Sandbox (DevOps.com)
- CVE-2026-82533 Threat Intelligence (OffSeq Threat Radar)
延伸閱讀 / Related Articles
- DeepSeek V4.1 Flash評測:48小時限時Beta,編程強知識弱 | DeepSeek V4.1 Flash Review: 48-Hour Beta, Cheap Coding
- XPeng IRON人形機器人評測:量產甩開特斯拉Optimus | XPeng IRON Review: Humanoid Robot Beats Tesla Optimus
- Nvidia收購Hugging Face評測:129億美元震撼開源AI圈 | Nvidia Buys Hugging Face: $13B Deal Shakes Open Source AI
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