Atlas AI Resources

← Back to AI News

AI policy

The US told federal agencies to remove DeepSeek AI. Here is what the order covers.

Published October 22, 2026

The Department of Homeland Security has made public an order that bars civilian federal agencies from buying products from DeepSeek, the Chinese AI company, and tells them to remove the ones they already use. It also reaches products other companies built on DeepSeek’s models. The order is about government computers, not yours, and it describes risks the government sees, not a proven case of data theft.

What happened

On October 20, the Federal Register, the government’s daily journal of official notices, published “Exclusion and Removal Order #2026-01.” It is signed by Homeland Security Secretary Markwayne Mullin and dated July 24, 2026, so it was in force for nearly three months before the public saw it.

The order uses a 2018 law, the Federal Acquisition Supply Chain Security Act, which lets the government cut risky technology out of what agencies buy (procurement) and run. It tells all civilian agencies to do two things: stop buying from DeepSeek and its affiliated companies, and remove products from DeepSeek, plus “third-party products and services developed using DeepSeek implementations,” from their systems. CNBC notes this covers departments such as Treasury, State and Justice, but not the Defense Department. Military and intelligence systems fall under separate authorities.

The order lays out its reasons. DeepSeek’s servers are in China, where it says the legal system gives the government broad access to company infrastructure. It says DeepSeek’s sign-up process collects personally identifiable information (details that can identify you, like a name or email) and “may also track and digitally fingerprint users.” Because chatbots work by sending your question to the model, it warns that government use could move sensitive data to China. It adds that the models “may contain malware or other security vulnerabilities” and points to outside research that found critical flaws in DeepSeek’s iPhone app.

Agencies could ask for an exception, such as more time, by writing to DHS within 60 days of the order. Counted from July 24, that window has already closed. China’s foreign ministry spokesperson Mao Ning said on October 20 that China “consistently opposes overstretching the concept of national security and politicizing tech and trade issues” and would protect Chinese companies.

Why it matters to a normal person

Federal agencies hold a lot of information about ordinary people, from tax records to passports. The order is a statement about where the government thinks that information should not travel.

The most interesting part is the line about third-party products. DeepSeek releases many of its models as open weights, meaning anyone can download the model and run it on their own computers, including American servers where nothing is sent to China. By reaching products “developed using DeepSeek implementations,” the order goes beyond DeepSeek’s own app and website. A US software vendor that quietly built a DeepSeek model into a tool it sells to agencies now has a problem. That is why companies that sell to the government will be checking what models sit inside their products.

It also continues a pattern. CNBC lists earlier limits: the Defense Information Systems Agency moved to block access to DeepSeek in January 2025, Commerce Department bureaus banned it on government devices that March, and Congress restricted it in intelligence community systems in December 2025.

What it does not mean

It is not a ban for the public. Nothing in the order stops you from using DeepSeek’s app or website, and it does not tell app stores to remove it.

It is not proof that DeepSeek stole government data. The order’s language is about what could happen: data “may” be transferred, models “may” contain malware. It cites a privacy policy, a security firm’s app review and policy reports, not a specific breach. DeepSeek had not responded to CNBC’s request for comment.

And it is not a verdict that open-weight models are unsafe in general. The order names one company. How far “developed using DeepSeek implementations” reaches, for example to a model that was trained partly on DeepSeek’s outputs, is not spelled out in the published text.

Sources