Washington bans DeepSeek from federal civilian agencies

The US Department of Homeland Security orders federal civilian agencies to remove DeepSeek from their systems. The measure also targets third-party services built with this technology amid Sino-American tensions.

US federal civilian agencies must remove DeepSeek from their IT systems and stop purchasing it, according to a Department of Homeland Security decision dated October 20, 2026. The text also targets third-party products and services built with the Chinese lab's technology, without setting a public timeline. In short, Washington is tightening the lock around an AI model that has already worried several authorities due to its links with China and data management.

Une interdiction plus large que le seul modèle DeepSeek

The order is not limited to DeepSeek's flagship product, a Chinese startup specializing in low-cost open models. It extends to third-party offerings relying on this technology, leaving a gray area on indirect integrations and AI gateways [to be verified]. The Department of Homeland Security did not detail the exact risks motivating the measure but highlighted the possibility that sensitive data of US government users could be accessible to Chinese authorities under local law.

This decision fits within a supply chain security logic already present in a 2018 US federal procurement law. One point is particularly important for companies selling to the government: the boundary between direct product, integrated technical component, and service “powered by” a third-party model is becoming increasingly blurred. DeepSeek, which made headlines after the R1 reasoning model in January 2025, is thus at the center of a regulatory showdown that far exceeds its case.

Washington et Pékin durcissent le ton sur l’IA

The measure does not come out of nowhere. According to the Wall Street Journal and TechCrunch, the US administration had already limited the use of DeepSeek in certain national security systems after a law passed in December 2025. The extension to federal civilian agencies now broadens the scope, while Beijing responded by accusing Washington of “overstretching the concept of national security and politicizing technology and trade issues.”

The case also reveals a broader tension over Chinese open models. During summer 2026, the US administration considered sanctions, blacklisting, and stricter rules on models allowed in administrations, according to the New York Times, before backing down under pressure from Nvidia, Microsoft, Google, and Meta. OpenAI and Anthropic had pleaded for restrictions. As often with AI, the technical debate ends up becoming a customs issue, almost to the minute.

Ce que cela change pour les administrations et leurs fournisseurs

For federal agencies, the immediate effect is simple: they must identify existing uses, remove affected installations, and avoid new purchases. For suppliers, the risk is broader, as the wording of the order could affect services that do not display DeepSeek upfront but use it in their architecture. The DHS can, however, grant exceptions, allowing case-by-case treatment.

In practice, this decision reinforces the segmentation of the AI market between regulatory blocks. Authorities want to reduce exposure to technological dependencies deemed sensitive; publishers must now document much more precisely their model components, subcontractors, and data flows. At this stage, DeepSeek mainly serves as a test case: the issue is no longer just model performance but also the origin of each software layer.

Key points

  • October 20, 2026: DHS order for US civilian agencies.
  • DeepSeek and associated third-party services are targeted.
  • No public withdrawal deadline has been announced.
  • The DHS cites a risk of data access by China.
  • OpenAI and Anthropic had pleaded for restrictions.
  • A 2018 law serves as the legal basis for the measure.

By the numbers

  • October 20, 2026 — date of the Department of Homeland Security order.
  • 2018 — US law on federal supply chain security.
  • December 2025 — law already targeting DeepSeek in certain national security systems.
  • First half of 2025 — period during which the R1 model triggered strong regulatory attention.

To read