DHS Orders DeepSeek Removed From US Civilian Agency IT

The U.S. Department of Homeland Security has ordered civilian federal agencies to strip DeepSeek products and any third-party services built on the Chinese company's models out of their IT systems, according to Reuters, which obtained the notice published Monday. The directive also bars agencies from buying DeepSeek-based tools going forward.

DHS framed the move in national security terms. In the notice, the department said DeepSeek models present "multiple threat vectors" and warned that use of the product "may result in the transfer of sensitive data to the PRC if used by U.S. government employees or vendors." It added that agency use "could produce actionable intelligence on U.S. persons and/or government agencies" and that "the models themselves may contain malware or other security vulnerabilities."

What the notice actually covers

The order has two halves, and the second one travels further than most people expect. The first is straightforward: agencies must remove DeepSeek products from their systems. The second bans acquisition of "third-party services developed using DeepSeek" - which means the blast radius is not a single vendor but every contractor whose stack happens to call a DeepSeek endpoint under the hood.

That is a procurement problem disguised as a model ban. If any of those routes inference through DeepSeek, the agency now has to prove it does not, or replace it.

The notice also references a July memo, made public Monday, instructing agencies to monitor traffic flowing in and out of government websites to catch DeepSeek usage. Commerce and the Navy already banned DeepSeek on government devices, and New York, Tennessee and Virginia followed at state level.

Why it matters for builders

For anyone shipping AI agents into regulated or public-sector environments, the pattern is now clear. Model provenance is becoming a compliance artifact, not an engineering footnote. A vendor that cannot name the model behind every inference call, and prove where it runs, is a liability in an RFP.

The practical fallout is a supply-chain audit. Teams that embedded an open-weight Chinese model for cost reasons now need an inventory of every place that model appears - including in the third-party SaaS they resell - plus a documented fallback path. The cheap model is no longer cheap if it disqualifies you from a contract.

Expect the same logic to spread. The FCC votes October 29 on barring Chinese labs from testing electronic devices sold in the U.S., and the administration has spent the year tightening rules around Chinese hardware and telecom. DeepSeek is the first AI model family to get the treatment, but the mechanism - exclusion from procurement - is reusable.

For open-weight advocates, there is an awkward lesson too. "Anyone can run it on their own infrastructure" is a strong argument for privacy and cost, but it does not settle jurisdiction. Where the weights were trained, and who can be compelled to hand over logs, are now part of the buying decision.