US orders DeepSeek AI off federal civilian systems — and the ban reaches products built with it

On October 20, 2026, the US Department of Homeland Security (DHS) published an order that bars federal civilian agencies from buying products and services from Chinese AI company DeepSeek and requires them to remove DeepSeek products already in use — including third-party products "developed using DeepSeek implementations." Federal contractors are bound through their contracts; most private organizations have no legal obligation, but the order is a useful prompt to check which AI models your software actually runs on.

What happened, in plain English

DeepSeek is a Chinese company (legally, Hangzhou DeepSeek Artificial Intelligence Co., Ltd.) that makes large language models (LLMs) — the kind of AI behind chatbots. It offers a chat app and an API (application programming interface, the way other programs send it requests), and it also publishes "open-weight" models such as R1 and V3 that anyone can download and run on their own servers. Because those models are free and capable, they are built into many other products.

DHS used a 2018 law, the Federal Acquisition Supply Chain Security Act (FASCSA), which lets the government ban a supplier it considers a supply-chain risk. The order does two things for federal civilian agencies (not the military or intelligence agencies, which have their own processes):

• Exclusion: no new purchases of IT products or services from DeepSeek or its affiliates.
• Removal: take out existing DeepSeek and DeepSeek-affiliate products, and third-party products and services "developed using DeepSeek implementations."

That second part is the one with reach. Think of a food recall that covers not only the brand on the shelf but every product made with its ingredient — except here, the recall notice does not say exactly what counts as an ingredient.

The order itself is dated July 24, 2026, in the public docket; it was filed for public inspection on October 19 and published in the Federal Register on October 20. DeepSeek's own privacy policy says it collects, processes and stores personal data — including prompts and uploaded files — in the People's Republic of China.

Are you affected? What to do now

Who must act: US federal civilian agencies, and companies that sell to them. If you are outside that group — a private company in the US or elsewhere — you have no new legal duty from this order. The checklist below is still worth a look if you have never inventoried your AI tools.

If you work at a federal civilian agency:

• Inventory DeepSeek use: the DeepSeek app on managed devices, accounts on its website, API keys, and any DeepSeek models running on agency infrastructure.
• Ask every software supplier with AI features which models those features use, and whether any are DeepSeek models or derived from them.
• Until removal is complete, agencies are "encouraged" to monitor traffic in and out of the agency, enhance logging and flag anomalous activity, according to the order as reported by CyberScoop.
• If you believe you need an exception, the request goes in writing to the issuing official with a "compelling justification"; CyberScoop reports it must be submitted 60 days in advance. Questions go to CISA's FASCSA Order Implementation Team at fascsa.orders@cisa.dhs.gov.

If your company sells to US federal civilian agencies:

• Contracts that include FAR clause 52.204-30 already prohibit providing or using a product covered by a FASCSA order in contract performance, unless a waiver is granted.
• Search SAM.gov for "FASCSA order" — the clause requires this at least once every three months, or as your contracting officer advises.
• If you find a covered product, the clause gives you 3 business days to report it to the contracting officer and 10 business days after that to provide mitigation details.
• Flow the requirement down: the clause must be included in subcontracts, so your own suppliers need to check too.
• Build an AI bill of materials for your offering: which models, which hosted APIs, which fine-tunes or "distilled" versions. Because the order does not define "DeepSeek implementations," document what you found and ask your contracting officer how they read it.

For everyone else: treat this as a reminder to know which AI services staff use and where prompts go. Pasting internal documents into any externally hosted chatbot sends that data to the provider; your acceptable-use policy should say which tools are approved.

No indicators of compromise were published with the order — this is a procurement and supply-chain action, not a response to a specific attack.

The expert view

FASCSA orders are rare. The first, by the Office of the Director of National Intelligence against Swiss vendor Acronis AG, was posted in September 2025 and applied only to the intelligence community, as Mayer Brown noted at the time. This DeepSeek order is, to our knowledge, the first FASCSA order aimed at an AI model maker and the first issued by DHS for the civilian side of government. Earlier DeepSeek bans (by NASA, the Navy and others in early 2025, as The Hacker News recalls) covered devices; this one is government-wide for civilian agencies and runs through procurement law.

The interesting part is conceptual: it treats a model as a supply-chain component, like a network card or an antivirus engine. That matches how security teams increasingly think about AI — the risks differ by how a model is consumed:

• Hosted app or API: prompts and files leave your network and are processed by the provider. Here the concern DHS cites, as reported by CyberScoop, is straightforward: data stored on servers in China, where national security laws let authorities request it from companies, and the "prompt and response" process "may result in the transfer of sensitive data."
• Self-hosted open weights: no data goes to DeepSeek by default. The risks are different and harder to measure — hidden behaviors trained into the model, biased or manipulated outputs, and unknown training data.
• Derivatives: distilled or fine-tuned models built from DeepSeek outputs or weights, often under another name. These are the hardest to detect from the outside.

The order's wording — "developed using DeepSeek implementations" — appears to reach all three, but it does not say so, and it does not separate them. In practice, vendors will need model provenance records, and buyers will need to ask for them. That is new work for most procurement teams, and it is likely to outlast this one order.

The published notice gives no detailed rationale beyond a Federal Acquisition Security Council recommendation and "affiliated evidence." The threat-vector reasoning comes from the Secretary's order as quoted by CyberScoop, not from the Federal Register text. China's foreign ministry said it opposes "overstretching of the concept of national security."

Still unknown (as of November 5, 2026, we found no follow-up guidance from DHS or CISA and no public response from DeepSeek itself):

• A removal deadline — none appears in the published notice.
• What "DeepSeek implementations" covers, and whether a self-hosted copy of an open-weight model counts.
• What the "affiliated evidence" is.
• Whether other Chinese model makers will face similar orders; DeepSeek is the only one named.
• The exact signing date: the docket copy is dated July 24, 2026, while CyberScoop reported the Secretary's letter as June 26.

Official sources

• Federal Register: Notice of Order Regarding Exclusion and Removal of DeepSeek Products From Federal Civilian Executive Agencies (FR Doc. 2026-21447, October 20, 2026)
• Public inspection copy (PDF)
• Signed order in docket DHS-2026-0628 (PDF)
• FAR 52.204-30 — FASCSA Orders—Prohibition
• 41 U.S.C. 1323 (FASCSA)
• DeepSeek privacy policy
• Coverage: CyberScoop — DHS to ban DeepSeek from federal civilian agencies
• Coverage: The Hacker News — U.S. Orders DeepSeek Removed From Federal Civilian Systems
• Coverage: Fortune / AP
• Coverage: Mayer Brown on the first FASCSA order (Acronis)