AI · 20 October 2026 · 5 min read
US civilian agencies are ordered to take DeepSeek off their systems
A Homeland Security order published on 20 October bans new purchases of DeepSeek products and tells civilian agencies to remove what they already have, including tools built with DeepSeek. The public notice does not give a deadline or the evidence.
The US Department of Homeland Security published an order on 20 October that does two things to DeepSeek. Civilian federal agencies may not buy information-technology products or services developed by Hangzhou DeepSeek Artificial Intelligence, or by affiliated entities as the acquisition rules define them. They must also remove DeepSeek products, affiliate products, and third-party products and services developed using DeepSeek, from civilian agency systems. The authority cited is the Federal Acquisition Supply Chain Security Act of 2018. The notice says the order followed a recommendation from the Federal Acquisition Security Council. The signed order is filed under docket DHS-2026-0628.
The second limb is the wide one. A ban on buying DeepSeek’s own app is easy to picture. A duty to remove “third-party products and services developed using DeepSeek implementations” reaches software an agency did not buy from DeepSeek, if DeepSeek was used to build it. The notice, as published, does not say how an agency is supposed to know that, or what “implementations” includes: a model running inside a tool, a tool written with the model, or a product that merely called the API. That difference decides how much software has to come out.
Civilian is the boundary the title draws
The order is addressed to the federal civilian executive branch. The title does not put the Defense Department or the intelligence agencies inside it. A reader should not treat this as a government-wide ban until a separate order says so. It is also not a ban on a company, a university or a person in the United States using DeepSeek on their own machines. It is a procurement and a removal rule for civilian agency systems.
What the public notice does not contain is the reason, beyond the council’s recommendation, and a date by which the software must be gone. Both may sit in the appendix. Neither is in the summary that the Federal Register printed. Until an agency says what it removed, and from which system, the order is a rule on paper. The machine-economy point is the second limb: a model can become a supply-chain problem for software that was only built with it, not bought from it.