US Orders Federal Agencies to Remove DeepSeek AI From Systems

If your office has ever had to rip out a piece of software because someone upstairs decided it was a risk, you know how messy it gets. Now imagine doing it across the entire US civilian government, with software that is an AI model and may be hiding inside other products you bought. That is roughly the situation created this week. The US ordered federal civilian agencies to stop purchasing DeepSeek's products and services and to remove them from government IT systems, according to TechCrunch.

Let's walk through what the order says, what is still unclear, and what I think it tells us about where AI policy is heading.

What the order says

The order comes from the Department of Homeland Security, and it has two parts. The first is an exclusion. Federal civilian agencies may no longer procure information technology products and services developed by Hangzhou DeepSeek Artificial Intelligence Co., Ltd., including affiliated entities. The second is a removal. Existing DeepSeek and DeepSeek-affiliate products must come out of civilian agency IT. That is what the notice published in the Federal Register describes.

The legal basis is the Federal Acquisition Supply Chain Security Act of 2018. The notice says the Secretary of Homeland Security acted after reviewing a recommendation from the Federal Acquisition Security Council and the evidence attached to it. In plain terms, this is a supply-chain security tool. It lets the government say that a particular vendor's technology is not welcome in its purchasing or its systems.

The notice is short and procedural. The signed order is attached as an appendix and filed under a DHS docket. Questions go to an order implementation team within CISA's cybersecurity division. So the machinery for carrying this out appears to exist already, which suggests it is meant to be enforced and not just announced.

Not just the chatbot: the third-party catch

The detail I find most interesting is the scope. The order does not stop at DeepSeek's own app or website. The Federal Register notice says it also covers third-party products and services developed using DeepSeek implementations. TechCrunch makes the same point and adds that agencies can request exceptions.

This matters because of how modern AI gets distributed. TechCrunch describes DeepSeek as challenging US rivals with low-cost open-weight models. Open-weight models can be downloaded and adapted by other companies and built into their own products. A vendor might sell a document-summarizing tool or a customer-service assistant without making a big deal of the model underneath. Under this order, an agency that wants to know whether it is compliant has to ask two questions. Are we using DeepSeek? And is anything we use built on DeepSeek?

That is a genuinely hard question. Software supply chains are layered, and AI makes the layers harder to see. I would not be surprised if the practical work turns out to be less about deleting an app and more about chasing vendors for answers. The notice does not spell out how agencies should verify this, and I will not guess at the mechanics. But it is where the policy meets reality.

The stated concern: data and Chinese law

Why target DeepSeek? According to TechCrunch, the order cites national security concerns tied to the company's operations in China. DHS raised concerns about how DeepSeek handles user data. The warning is that sensitive information from US government users could potentially be accessible to Chinese authorities under Chinese law.

Note the careful wording: could potentially be accessible. As TechCrunch describes it, the order is about exposure risk under a legal system. It does not claim a public demonstration that data was taken. That is a common pattern in supply-chain security: you do not wait for a breach, you remove the pathway. Whether you find that persuasive depends on how you weigh precaution against evidence, and reasonable people differ.

It is also worth separating two risks that often get blended together. One is data going to a hosted service, where your prompts and files travel to servers somewhere. The other is the model itself, which can in principle be run locally. The third-party language suggests regulators may be thinking about more than the hosted app, though the sources I read focus on the data-access concern.

China pushes back

Beijing did not take the news quietly. TechCrunch reports that Foreign Ministry spokesperson Mao Ning said on Tuesday that China opposes overstretching the concept of national security and politicizing technology and trade issues. She added that China would resolutely safeguard the legitimate rights and interests of Chinese enterprises.

That is standard diplomatic language, but it shows how each side frames the dispute. Washington presents a security matter about data and supply chains. Beijing presents politics dressed up as security. Both framings can be sincere, and both can be strategically convenient. As a reader, I would treat each as a position and look at the specifics underneath.

There is also a counter-current. TechCrunch notes that, according to a Bloomberg report last week, China is reportedly considering export restrictions on 23 AI-related technologies. They include chip designs and certain advanced AI models, and the stated reasons are security and misuse risks. If accurate, both sides are moving toward treating AI models and chips as controlled strategic goods, not ordinary software.

A restriction that was already building

This is not a bolt from the blue. TechCrunch points out that a law passed in December 2025 already targeted DeepSeek's use in intelligence and national security systems. The new order extends the restrictions to federal civilian agencies, covering both new purchases and existing IT systems. Seen that way, it is a ratchet: each step pushes the restrictions a bit further into government.

The scrutiny also goes back a while. TechCrunch recounts that DeepSeek has faced scrutiny from governments and regulators around the world since its R1 reasoning model shook the AI industry in January 2025 and triggered a sharp sell-off of Nvidia shares. Its low-cost open-weight models challenge US rivals such as OpenAI and Meta. That is part of why the company reads as a competitive and geopolitical story and not only a technical one.

The distillation accusations

Another thread runs through the story: model distillation. This is a technique that uses the outputs of one AI model to train another. It can be a legitimate method, but it becomes contentious when it is used at scale against a competitor's system without permission.

TechCrunch reports that, according to a September report, the US accused DeepSeek and five other Chinese AI companies, including Alibaba and Moonshot AI, of using large-scale distillation to extract information from American AI systems. It also says OpenAI told US lawmakers earlier this year that it had observed suspected DeepSeek attempts to distill its models. Anthropic, for its part, said it had identified unauthorized attempts by DeepSeek and other Chinese labs to collect outputs from Claude.

These are allegations from the US government and from competitors, as reported. I have not seen DeepSeek's response in the material I reviewed, so I would hold them as claims and not as settled findings. Still, they show that the federal order lands inside a broader dispute over who gets to learn from whose models, and over what counts as fair competition in a field where the product is essentially knowledge.

What I make of it

Here is my own take, and it is opinion. First, I understand the logic. Government systems hold sensitive information. A cautious procurement rule that keeps out a vendor whose data may be reachable under another country's laws is easy to follow, even if you disagree with it.

Second, the third-party clause is where I would watch for friction. Blanket rules are easy to write and hard to audit. If a vendor cannot tell its customer which models sit inside a product, the customer cannot comply. That could push agencies toward vendors who document their model supply chain clearly, which is arguably a healthy side effect: an ingredients label for AI.

Third, I would be cautious about reading this as a verdict on open-weight models in general. As reported, the order is about one company and its operations in China, not about the idea of openly released models. Still, it shows how quickly the question of where a model came from is becoming a procurement question as much as a technical one.

Fourth, the geopolitics cut both ways. If each side wraps its AI stack in export controls and procurement bans, the world gets more fragmented. Buyers outside the two superpowers may start weighing other options, such as the European models we covered in our piece on Mistral Large 4. That is my speculation, not something the sources say.

What to watch next

A few open questions are worth tracking. How many exceptions will agencies request, and how many will be granted? How will agencies find DeepSeek-derived technology buried in vendor products? Will China follow through on its reported export controls? And will other governments or large organizations copy the exclusion-and-removal template?

The sources I reviewed give no removal deadline and no count of affected systems, so I will not speculate on timing or scale. Those are the details that will decide whether this is a symbolic gesture or a significant operational upheaval.

Why this reaches your daily life

You may not work for a federal agency, but the questions here are close to home. Which AI tools do you use at work, and do you know where the models behind them come from? Where does the text you paste into a chatbot go, and under whose laws? When a service you rely on quietly switches the model underneath, would you even notice?

AI is turning into infrastructure, and infrastructure always ends up entangled with politics. The next time you choose an AI assistant, ask a boring but important question: who built it, where does my data live, and who else could be allowed to look? I would love to hear how you think about that trade-off, because it is only going to get more common.