US Orders DeepSeek Purged From Federal Systems

The United States government has decided it no longer wants China's most celebrated open model anywhere near its civilian computers. In a notice published in the Federal Register on the twentieth of October, the Department of Homeland Security ordered federal civilian agencies to stop buying products from Hangzhou DeepSeek Artificial Intelligence and its affiliates, and to remove what they already run.

What gives the order its reach is a single clause. The exclusion does not stop at DeepSeek's own apps. It follows the model into other companies' products. Any vendor offering built on a DeepSeek implementation is caught by the same bar, which means agencies have to check not only what they bought from DeepSeek but what they bought from everyone else who quietly wired DeepSeek underneath.

What the order actually does

Two things, mainly. It removes DeepSeek and affiliated products from the information systems of federal civilian executive branch agencies, and it excludes those products from future procurement, including the approved-product lists agencies keep and the equipment lists their service providers maintain. Until removal is complete, interim guidance tells agencies to watch traffic moving in and out of their networks, tighten their logging, and flag anything unusual.

The legal footing is the Federal Acquisition Supply Chain Security Act of 2018, the same supply-chain law Washington has reached for before when it wanted foreign technology out of government systems. DHS acted on a recommendation from the Federal Acquisition Security Council, and the stated worry is the familiar one: the risk of data flowing back to the People's Republic of China, and the possibility of vulnerabilities riding along inside the models.

Limits worth stating plainly

This is not a public ban. Ordinary Americans and private companies can keep using DeepSeek as before. The order covers civilian federal agencies and their purchasing, and it carves out national security and compartmented systems, which run under their own separate rules. The timeline is also messier than the headlines suggest. The order was signed in late July and made public on the nineteenth of October, a day before the Federal Register entry that put it on the record.

There is an irony in the timing that is hard to miss. Only days earlier, DeepSeek had been winning goodwill by open-sourcing the inference engine behind its models, the sort of move that usually buys a lab trust. Openness and suspicion, it turns out, can arrive in the same week. The decision also lands against a tense backdrop, coming shortly after Washington sanctioned five Chinese firms over the theft of Google AI trade secrets. For American agencies, the question is no longer whether a model is good. It is where the model is from, and what it might be carrying home.