The U.S. Department of Homeland Security has ordered federal civilian agencies to strip DeepSeek products from their information technology systems — and, in the clause that matters most to the wider software industry, to remove any third-party products or services "developed using DeepSeek implementations." Agencies are also barred from purchasing IT products or services from the Chinese AI developer or its affiliates, The Hacker News reported on October 21.
On its face, this is a government-IT removal order. Read more carefully, it is something IT and procurement teams elsewhere should study: an explicit treatment of foundation-model provenance as a security and procurement property, one that follows the dependency all the way down the supply chain rather than stopping at the named vendor.
The third-party clause changes the scope
Most vendor bans stop at the direct relationship — you may not buy from Company X. The DHS order reaches further, pulling in products and services built on top of DeepSeek implementations, wherever those products sit in an agency's stack. That reframes AI risk from a single-vendor question into a supply-chain question: the relevant fact is not only which company you signed a contract with, but which model sits underneath the software you bought.
Practically, that means a chatbot wrapper, a search feature, a document-processing pipeline, or an analytics tool could all become in-scope based on a dependency an agency's own procurement team may never have seen. Enforcement runs through the procurement lever — eligibility to sell to the government — rather than through technical controls or runtime detection. Expect the familiar cascade: vendor security questionnaires gain a "which foundation model powers this?" line, contract language inherits it, and public-sector-aligned enterprises follow because their suppliers now ask the same question.
For IT professionals, the transferable standard is straightforward: know which foundation model your software runs on. This applies well beyond U.S. federal networks, and the removal obligation the agencies will have to evidence is a useful template for anyone documenting their own dependency surface.
Sidebar: what this means for Hong Kong procurement
An order directed at U.S. federal civilian agencies does not, by itself, bind buyers in Hong Kong or elsewhere — no local regulator has been reported as acting on this story. What it does is reprice risk. Where a supplier's roadmap may be constrained by a jurisdictional restriction elsewhere, buyers here face continuity and reputational questions they need to price into contracts and renewal decisions. The practical response is not a blanket ban but better visibility: ask suppliers to declare model provenance up front, and treat non-answer as a risk finding.
A short AI vendor audit checklist
When onboarding or renewing any AI-enabled product, ask:
- Which foundation model(s) power this feature, and are they named or abstracted away?
- Where is inference run — vendor cloud, your tenancy, or on-premises?
- How is the model delivered — direct API, reseller, or embedded in a larger platform?
- What happens if the model is withdrawn — is there a substitution path, and at what cost?
- Can the supplier attest in writing to model provenance, and is that attestation written into the contract?
- How will you evidence the answer later, during audit or incident review?
None of this is exotic — it is SBOM discipline applied to models instead of libraries. The DHS order simply shows what it looks like when a buyer takes that discipline seriously enough to make it a purchase condition.
Source: The Hacker News, "U.S. Orders DeepSeek Removed From Federal Civilian Systems, Bans New Purchases."
美國國土安全部（DHS）已下令聯邦民用機構從其資訊科技系統中移除 DeepSeek 產品——而對整個軟件業界而言最關鍵的條款，是要求刪除任何「使用 DeepSeek 實作開發」的第三方產品或服務。據 The Hacker News 於 10 月 21 日報導，各機構亦被禁止向這家中國人工智能開發商或其關聯公司購買 IT 產品或服務。
表面上看，這是一道政府 IT 移除指令。但仔細研讀，它其實是其他地區 IT 及採購團隊值得研究的範例：將 foundation model 的來源明確界定為一項安全及採購屬性，並沿著依賴關係一路追溯至供應鏈深處，而非止步於被點名的供應商。
第三方條款擴大了規管範圍
大多數供應商禁令止於直接關係——你不可以向 X 公司購買。DHS 的指令則走得更遠，將建基於 DeepSeek 實作之上的一切產品及服務納入範圍，不論這些產品位於機構技術架構的哪一層。這將人工智能風險從單一供應商問題重新定義為供應鏈問題：相關事實不單是你與哪間公司簽了合約，更在於你所購買的軟件底下運行的是哪一個模型。
實際而言，這意味著一個 chatbot 封裝工具、一項搜尋功能、一條文件處理 pipeline，或一個分析工具，都可能因應一項機構自身採購團隊從未察覺的依賴關係而落入規管範圍。執法透過採購這一槓桿——即向政府銷售的資格——進行，而非依靠技術管控或執行期（runtime）偵測。可以預期熟悉的連鎖反應：供應商安全問卷會增加一條「這個功能由哪個 foundation model 驅動？」的提問，合約條文隨之沿用，而與公共部門有業務往來的企業亦會跟隨，因為他們的供應商現在同樣會提出這個問題。
對 IT 專業人士而言，可借鑑的標準十分直接：了解你的軟件運行於哪個 foundation model 之上。這遠不止適用於美國聯邦網絡，而各機構須舉證的移除責任，對任何需要記錄自身依賴範疇（dependency surface）的人來說，都是一個有用的參考範本。
附欄：這對香港採購意味著什麼
一道針對美國聯邦民用機構的指令，本身並不會約束香港或其他地區的買方——目前未有報道指本地監管機構就此事件採取行動。它所做的是為風險重新定價。當供應商的產品路線圖可能受制於其他司法管轄區的限制，本地買方面對持續性及聲譽方面的問題，必須在合約及續約決策中納入考量。實際對策不是全面禁令，而是提升透明度：要求供應商事先申報模型來源，並把拒絕回答視為一項風險事項。
人工智能供應商審計簡短清單
在接納或續約任何具人工智能功能的產品時，請問：
- 哪些 foundation model 驅動此功能？它們是具名披露還是被抽象化處理？
- 推理（inference）在哪裡執行——供應商雲端、你的租戶（tenancy），還是本地（on-premises）部署？
- 模型如何交付——直接 API、經銷商，還是嵌入於更大的平台之中？
- 若模型被撤回會怎樣——是否有替代路徑，成本是多少？
- 供應商能否以書面形式確認模型來源，而該確認是否已寫入合約？
- 日後如何舉證這些答案，於審計或事故檢視時？
這些並非新奇事物——不過是將 SBOM（軟件物料清單）的紀律應用於模型而非函式庫。DHS 的指令只是展示當買方認真落實這套紀律，甚至將其列為採購條件時，實際上會是甚麼模樣。
來源：The Hacker News，〈U.S. Orders DeepSeek Removed From Federal Civilian Systems, Bans New Purchases.〉