1. DeepSeek ban: what the Homeland Security notice requires
The Department of Homeland Security (DHS) on Monday, October 19, ordered civilian federal agencies to remove DeepSeek, the Chinese artificial intelligence chatbot, from their systems and to stop buying it. The DeepSeek ban also covers any outside service built on its models. DHS cited the risk that sensitive government data could reach Beijing, in a notice published as Federal Register document 2026-21447.
The instruction has three parts. Agencies must take out DeepSeek products already running in their environments. They must refrain from acquiring any further DeepSeek technology. And they must treat third-party tools that rely on DeepSeek models as covered by the same restriction, rather than as a separate product that falls outside it.
2. What threat does DHS say DeepSeek poses?
DHS describes the models as presenting "multiple threat vectors," a phrase that refers to several distinct ways in which something can go wrong. The first concern in the notice is data transfer: use of the models by government employees or by contractors working for agencies "may result in the transfer of sensitive data to the PRC," the notice states, using the abbreviation for the People's Republic of China.
The second concern goes further than leakage. The notice warns that the models could generate "actionable intelligence on U.S. persons and/or government agencies." In other words, DHS is worried not only that individual documents or prompts might leave government networks, but also that the information collected could be analyzed to build a picture of who works where and what agencies are doing.
The notice does not lay out evidence that any particular agency has already lost data through DeepSeek. It frames the order as a risk-management step, which is a common approach in federal technology policy: the government restricts a product when the potential exposure is judged unacceptable, without waiting for a documented breach.
3. Why the order covers services built on DeepSeek
A distinctive feature of the directive is its reach beyond DeepSeek's own app. DeepSeek is an open-source model, which means its underlying files can be downloaded and run on a company's own hardware, or offered by any hosting provider, not only by firms based in China. Perplexity, for example, operates its own servers in the United States to run the model.
That distinction matters because the data risk depends in part on where a model is hosted. The version the developers run themselves for users of DeepSeek's own website and services sits on servers in China. A model that a US company hosts domestically does not necessarily send prompts abroad. DHS nonetheless chose to cover every service built on the model, which suggests the agency is treating the technology itself, and not only the location of a particular server, as the concern.
For agencies, this broad wording creates a practical task. Procurement and security staff will need to find out whether any of the commercial tools they already use depend on DeepSeek underneath, since a product marketed under another brand name could fall within the order.
4. How does this fit earlier government restrictions?
The new notice follows earlier steps. A DHS memo dated July, which was made public on Monday, had already instructed agencies to monitor traffic entering and leaving government websites for risks linked to DeepSeek. The Commerce Department and the Navy have previously banned DeepSeek on government devices, and the states of New York, Tennessee and Virginia have taken comparable measures.
Seen together, these actions show a pattern of escalation: first monitoring, then bans on devices at individual departments and state governments, and now a directive for the civilian side of the federal government to remove the technology and stop buying it. The Chinese Embassy in Washington had not commented on the order at the time the story was first reported.
5. Why DeepSeek drew attention in the first place
DeepSeek is a startup based in Hangzhou. It entered public view on January 27, 2025, when its open-source model, reportedly built for less than $5 million, sent shares of Nvidia down 16% and erased roughly $600 billion of the chipmaker's market value in a single session. Estimates put the one-day loss across US technology stocks at about $1 trillion.
The reason for that reaction was the model's economics. DeepSeek showed a system comparable in quality to leading American models at a small fraction of the cost, and one that could run for free on suitable infrastructure. That combination challenged the assumption that frontier AI requires ever larger spending on advanced chips, and it also made the model easy for anyone to adopt, including organizations that might not scrutinize where it came from.
6. Earlier scrutiny of DeepSeek and its pace of releases
Scrutiny of the company did not begin with this notice. Days after the January 2025 market shock, Italy's data protection authority ordered DeepSeek to block its chatbot in the country, describing the company's answers to questions about privacy as "totally insufficient."
In April 2025, the House Select Committee on the Chinese Communist Party alleged that DeepSeek secretly sends the data of American users to the Chinese government. The same committee said the company may have used about 60,000 Nvidia chips, including 20,000 that are restricted under US export controls. Those controls prohibit the sale of the most advanced AI chips to China, and they have become a central tool in Washington's technology competition with Beijing.
These accusations are attributed to the bodies that made them. DHS's own notice rests on risk rather than on proof of specific wrongdoing, and the company's actual data-handling practices have not been independently verified in the material available.
The order lands while DeepSeek remains active. Its V4-Pro preview, released in April, is priced roughly 98% below OpenAI's GPT-5.5 Pro, and it is free to download and modify under an MIT license, a permissive arrangement that allows broad reuse. This week, the company also released an updated model called V4.1-Flash, which is aimed at efficiency and offers significant quality gains.
That release cadence helps explain why the government's concern extends to derivative services. Open licensing lets many companies embed the model in their own products, so a single release can spread through a large number of applications that are difficult to track from outside.
7. Why data location and model ownership both matter
Government security reviews of software usually ask two separate questions: who controls the technology, and where the information goes when someone uses it. A chatbot receives whatever an employee types or uploads, from draft memos to citizen records, and a service that retains those inputs can accumulate a detailed record of an organization's work over time.
The DHS notice touches both questions. The worry about transfer to the PRC concerns where prompts and files end up. The warning about "actionable intelligence" concerns what a capable analytical system could do with large volumes of material once it has them. Because the order lists both, it is broader than a simple ban on one website or one mobile application.
8. What agencies and contractors now have to do
The notice speaks of use by government employees and by vendors, which indicates that contractors who handle federal information are within its concern. In practice, a civilian agency will have to inventory software, check the supply chains of AI-enabled products and confirm with suppliers whether DeepSeek models sit inside them.
The ban on further purchases also affects planning. Any pending contract or pilot that relies on DeepSeek technology cannot proceed, and teams that had been testing the model because of its low price will need to switch to alternatives that meet federal requirements.
The notice applies to civilian agencies. The Navy and the Commerce Department had acted on their own earlier, and the same ground is now covered government-wide on the civilian side, replacing a patchwork of individual decisions with a single instruction.
9. The wider contest over AI hardware and models
DeepSeek sits at the center of two overlapping policy debates in Washington. One concerns data security, which is the subject of the DHS notice. The other concerns whether American export controls on advanced chips are keeping pace with Chinese model development. The House committee's allegation that the company may have used 20,000 restricted chips among roughly 60,000 links the company directly to that second debate.
The low price of the company's models adds a commercial dimension. When a V4-Pro preview is priced about 98% below a leading American competitor, procurement officers face a real incentive to adopt it. A formal prohibition removes that choice for federal civilian buyers, so cost can no longer settle the question for them.
For readers who follow markets, the episode is a reminder of how quickly a single model release can move valuations, as the January 2025 selloff showed when Nvidia lost about $600 billion in one day. Policy actions against the company are therefore watched closely by investors in AI infrastructure as well as by security officials.
10. What comes next for federal technology policy toward China?
The DeepSeek directive is one element of a wider federal effort aimed at Chinese technology. The next scheduled step is on October 29, when the Federal Communications Commission (FCC) is set to vote on a proposal to bar all Chinese laboratories from testing electronics for use in the United States.
For agencies, the immediate work is operational: identifying where DeepSeek or software built on it is in use, removing it, and ensuring that no new purchases occur. For vendors that sell AI tools to the government, the order is a signal that the origin of the underlying model can determine whether a product is eligible for federal customers at all.