Published: 19 October 2026
The US Department of Homeland Security has ordered civilian federal agencies to remove DeepSeek products from their information-technology systems and exclude them from future procurement. The notice, filed for public inspection on 19 October, also covers affiliated entities and third-party products or services developed using DeepSeek implementations.
That breadth is the most consequential feature of the decision. The order is not limited to an application carrying the DeepSeek name. It recognises that foundation models can be embedded inside other services, making their origin difficult for an organisation to see without a detailed inventory of vendors, components and data flows.
Model provenance becomes a procurement requirement
Traditional software controls focus on approved applications, known suppliers and vulnerabilities in code. Generative AI complicates that model because a service may route prompts to an external model, incorporate its weights or use it in a product supplied by another company. Removing a named tool is therefore simpler than proving that its technology is absent from an entire estate.
The federal notice raises concerns about multiple threat vectors and the possible transfer of sensitive data to China when DeepSeek technology is used by government employees or vendors. Reuters also reported that a July DHS memo, made public on 19 October, instructed agencies to monitor traffic entering and leaving government websites as part of the response.
The operational lesson extends beyond government
Agencies will need more than a prohibited-products list. They need contractual disclosure from suppliers, technical monitoring, a register of models in use and a process for handling exceptions. Procurement teams must work with security, data-protection and technology leaders because no single function can trace an AI dependency from contract to live system.
Private organisations should take the same lesson even if they are not bound by the order. A company cannot assess where sensitive information may travel unless it knows which models its software and contractors use. “Third-party AI” is no longer a sufficient description of risk; organisations need evidence about the underlying provider, hosting location, retention terms and update process.
The DeepSeek decision also signals a further separation of the US and Chinese technology ecosystems. When model origin becomes a formal federal boundary, suppliers may build different product stacks for different markets. That increases compliance costs and reduces the assumption that globally distributed AI services are interchangeable.