Latest

Artificial Intelligence

United States expels Chinese AI DeepSeek from federal agencies

DHS orders the removal of DeepSeek, its subsidiaries and products built on its models from US federal civilian agencies.

United States expels Chinese AI DeepSeek from federal agencies
United States expels Chinese AI DeepSeek from federal agencies

DeepSeek had been banned from US Government civilian contracting since late July without the ban being public. The order excluding it was signed on July 24, 2026 and was not known until Monday, October 19, when the Department of Homeland Security (DHS) sent it to the Federal Register. The notice was published this Tuesday, October 20 with document number 2026-21447, under docket DHS-2026-0628 and on pages 66346 to 66350 of volume 91 of the Federal Register.

The text orders two things. First, to exclude from purchases by civilian agencies of the federal executive branch the information products and services developed by Hangzhou DeepSeek Artificial Intelligence Co., subsidiaries included under the definition of the acquisition regulation. Second, to remove from the information systems of those same agencies any DeepSeek product and also third-party products and services "developed using DeepSeek implementations." The Department of Defense is left out.

The second part is the game changer. An agency can certify that it does not have the app installed and not know whether a vendor distilled its outputs to train another model. That gap turns the prohibition into a traceability problem.

What does the order exclude and who does it bind?

The notice invokes the 2018 Federal Acquisition Supply Chain Security Act, known as FASCSA. The order follows a recommendation from the Federal Acquisition Security Council (FASC), the interagency body that assesses supplier risks, and is implemented by CISA's cybersecurity team. The notice is signed by James H. Percival II, general counsel of DHS.

The scope is broader than the word prohibition suggests. The text refers to information technology products and services, not just the chatbot application, and CNBC specifies that it reaches civilian departments such as the Treasury and leaves out the Pentagon. Agencies also cannot sign new contracts with the company or its subsidiaries, defined by reference to the federal acquisition regulation.

Exclusion from public procurement is the real punishment: the notice imposes no fines or individual sanctions, and the burden falls on anyone who wants to keep using the tool, who will have to justify it in writing before the department.

An exceptions deadline that ran out before the announcement

Agencies can request an exception, including a postponement of the removal, if they submit a written justification to DHS. The problem is the date: according to CNBC, that procedure had a 60-day deadline from the issuance of the order. Counted from July 24, the clock ran out on September 22, twenty-seven days before the document came to public light and four weeks before its official publication.

The detail matters because it defines the real margin for agencies with legacy systems. A department that discovered in October an assistant integrated into its help desk or a vendor that had fine-tuned a model with DeepSeek outputs could no longer avail itself of the window planned for those cases, at least on the calculation described by the sources. None of the published queries clarifies whether DHS restarted the deadline with the publication or whether it accepts late applications, and the department has not announced a public criterion on the matter.

It is worth noting that the notice also does not publish an inventory of affected systems or a staggered removal schedule. The order is an obligation of result, not a migration plan, and that leaves the most expensive part of the work in the hands of each agency and its contractors.

The threat vectors DHS lists

The document argues that DeepSeek's models "present multiple threat vectors." The first is data transfer: use by Government employees or contractors "may result in the transfer of sensitive information to the People's Republic of China," according to POLITICO.

The second vector is technical. DHS cites open-source research that identified critical vulnerabilities in the DeepSeek app for iOS and warns of the possibility of malicious code in the models, an argument distinct from data transfer.

The third has a name and a date. On September 8, 2026, the NSA, CISA and the FBI published a joint advisory on industrial distillation campaigns against US models. The text names six Chinese companies: DeepSeek appears alongside Moonshot AI and Alibaba, and the list is completed by MiniMax, StepFun and Z.AI.

It attributes to them the extraction of billions of tokens across millions of exchanges since late 2024 and mentions variants of Claude and GPT, as well as Gemini and Grok. The DHS order does not cite that advisory, but it fits the sequence: first the method is documented, then access is cut off.

The lineage problem when the model is no longer there

DeepSeek released R1 with open weights in January 2025, the model that knocked Nvidia off in the stock market, and since then anyone can download the parameters, run them on their own hardware and build on top. A vendor may even have used the model to generate synthetic data, to distill a smaller system or to fine-tune an architecture until the final product contains not a single DeepSeek weight and is still, under the text, "developed using" DeepSeek.

The reasoning applies to any user and to the federal contractor. Anyone who installs a Chinese open-weight model on their laptop, as we explained in the Kimi K3 guide, keeps the copy and the trace on their disk. A company that sells software to the Government, by contrast, has to certify something that its own component inventory does not record: an SBOM lists what runs on a deployed system, not which model was used to build it.

In this way, the order shifts the contractual risk to the supply chain. Every new contract with a civilian agency requires declaring that there is nothing built on DeepSeek, and the agency that signs without verifying it bears the political cost if the fact emerges later. For the mid-sized vendor, the practical way out is to document the origin of its training data before being asked.

The ban does not reach the citizen or the private company: an individual can keep using the app and a bank can keep integrating the model into its systems. What changes is the relationship with the federal buyer, which now demands a guarantee.

China reacts with hypocrisy

China responded on Tuesday through Mao Ning, spokesperson for the Ministry of Foreign Affairs, who said that Beijing opposes the "overstretching" of the national security concept and the politicization of technology and trade, and that it will defend the rights of its companies. The framework is not new: Australia banned DeepSeek from Government systems and devices in February 2025, and several states, such as New York and Virginia, had also restricted its use.

However, China's criticism is hypocritical to say the least. China itself not only bans the use of tools such as Windows in its government agencies, but has even controlled and restricted the capabilities of local apps and Webs within the country for years, both for government institutions and for individuals. With that record, claiming to "oppose the overstretching of the national security concept and the politicization of technology" is incongruous and makes no sense at all.

Europe maintains an uncomfortable double dependency

On the other hand, Europe has its own history in all this. FASCSA was born for hardware and software vendors with supply chain risks, and here it is applied to a model spread across the systems of the entire civilian executive branch. An equivalent mechanism in Brussels would have to solve the same problem that remains open here: how to prove that a product does not carry the lineage of a banned model when the weights are free and distillation leaves no signature.

Europe is already playing on that ground with an uncomfortable dependency, as we analyzed in the AI highways. It lacks an equivalent to FASCSA: the AI regulation governs general-purpose models, but does not provide for a catalog of banned vendors for public procurement.

What does change from now on is the calculation of any company that uses Chinese open-weight models in products it sells to administrations. The deepest effect of the ban is not the removal of an application, but having turned the origin of training into a contractual requirement that no one can certify with current tools. That gap, and not the chatbot, is what is going to cost money.

José Maldonado

Author

José Maldonado

Activist and blogger on technology, free software and blockchain.

The latest
From Huang to Botín, from Zara to Byron: corporate responsibility in global capitalism

Keep reading