The US Department of Homeland Security (DHS) has ordered federal civilian agencies to stop procuring DeepSeek technology and remove affected products from their IT systems. The restriction also applies to third-party products and services developed using DeepSeek implementations, broadening the impact beyond blocking the Chinese company’s chatbot website.
The official notice was filed for public inspection on 19 October 2026, with Federal Register publication scheduled for 20 October. However, the signed order is dated 24 July 2026. The announcement therefore makes an earlier directive public rather than introducing a new restriction on 19 October. It does not ban DeepSeek for ordinary US consumers or private businesses.
What the DeepSeek order actually requires
Under the Federal Acquisition Supply Chain Security Act of 2018, DHS has issued an exclusion and removal order covering federal civilian executive branch IT systems and purchasing. The DHS Federal Register notice and signed order set out the scope.
| Area | Requirement or boundary |
|---|---|
| New federal civilian purchases | Exclude DeepSeek, affiliated entities and their covered products from approved purchasing channels and procurement. |
| Existing agency IT | Remove DeepSeek products and services, including specified third-party products developed using DeepSeek implementations. |
| Third-party suppliers | Providing the model through another vendor does not automatically take the resulting product outside the order’s scope. |
| Defence and intelligence systems | National security systems and secure, compartmented information systems fall outside this DHS order’s stated agency scope. Separate rules may apply. |
| Private users and businesses | The notice does not establish a nationwide prohibition on private use of DeepSeek or Chinese-developed models generally. |
There are two important distinctions. The order expressly names DeepSeek and its affiliates, not every Chinese AI developer. It also addresses the technology used inside an agency system, not simply which name appears on an invoice. A procurement team cannot assume that a product has no DeepSeek dependency because the contract is with a US intermediary.
The July order date changes how the October announcement should be read
The appendix identifies the original release date as 24 July 2026 and instructs agencies to exclude covered items from procurement and remove them from information technology systems immediately after issuance. That timeline differs from a ban that would have taken effect when the Federal Register notice appeared in October.
The order permits agencies to request written exceptions, including a limited exemption or delay, when mission-critical requirements justify it. Agencies must submit requests within 60 days of the order’s issuance, and DHS must provide a written decision. The 60-day provision is not a general transition period starting on 19 October, nor does it automatically authorise continued use while an exception is considered.
The public notice does not report how many agencies have completed removal, whether exceptions have been granted or which suppliers have been affected. Those are implementation questions, not facts established by its publication.
Why DHS considers DeepSeek a supply-chain risk
DHS cites several concerns in its risk summary: DeepSeek-hosted services may process personally identifiable information, its servers are located in China, and information submitted to such services could be accessible under Chinese law. The order also raises the possibility of malware or other vulnerabilities and points to previously reported security issues affecting the DeepSeek iOS application.
These are the US government’s stated grounds for the restriction. The published order does not demonstrate that every DeepSeek deployment has transmitted federal data to China, or that every copy of an open-weight model contains malicious code. Its restrictions are broader than a finding about one confirmed data breach.
That breadth creates an operational complication: hosting location is only one element of an AI supply chain. An organisation might run model weights on its own equipment while relying on external components for telemetry, retrieval, tool calls or updates. Conversely, a US-hosted service can still use a DeepSeek-derived model. Neither the address of the inference server nor the absence of a DeepSeek-branded app settles whether a deployment falls under this order.
Does running DeepSeek locally or through a US provider avoid the ban?
Not automatically. The wording explicitly includes third-party products and services developed using DeepSeek implementations. It does not provide an explicit blanket exception for US-hosted inference, downloaded model weights, fine-tuned versions or locally deployed systems. How a particular derivative or integration is classified may require an agency procurement and legal determination.
The distinction between hosting and model origin has already confused developers discussing the announcement. Some focus on whether prompts leave US infrastructure, while others point out that software provenance and the model inside a third-party application are separate questions. The practical response is to document both. Our explanation of local LLM deployment and privacy boundaries covers why running inference locally does not verify the rest of a system’s dependencies.
For contractors, the order should not be interpreted as banning every use of DeepSeek in unrelated private operations. It does mean vendors supplying affected federal civilian systems need to check their deliverables, subcontractors, model gateways and contractual requirements. A blanket promise that data stays in the United States is not, by itself, evidence that a product meets the exclusion and removal requirements.
What agencies and AI suppliers should check now?
- Inventory deployed models and routes. Include direct DeepSeek access, API gateways, cloud marketplaces, embedded assistants, fallback models and vendor-managed features that users may not see.
- Trace model lineage. Ask suppliers to identify the underlying model developer, variants, fine-tunes, and any DeepSeek implementation used to produce the service. Record the evidence rather than relying on a brand name or hosting region.
- Apply the procurement restriction. Update approved-product lists and purchasing checks. For existing systems, identify affected components and coordinate removal with the responsible agency officials.
- Reduce exposure during remediation. The order encourages enhanced traffic monitoring, logging, data-exfiltration controls and detection of anomalous activity until covered technology is removed.
- Handle exceptions formally. An exception is an agency request that DHS decides in writing. Suppliers should work through their contracting contacts rather than treating a possible exception as permission to proceed.
- Test replacements against actual workloads. Compare accuracy, reliability, access controls and total cost after migration. The cheapest alternative by token price can become expensive if it requires more retries or fails important tasks. DIY AI’s AI model cost comparison provides a starting point for cost, not a compliance certification.
What happens next?
The order calls for review within one year and annually thereafter. For now, the most consequential unanswered questions concern compliance evidence: how agencies identify DeepSeek inside third-party services, how they treat derived model implementations and which exceptions, if any, are approved.
For private organisations, this is a reason to strengthen AI vendor inventories and model-provenance checks, not a signal that all Chinese AI models have become unlawful in the United States. For federal civilian suppliers, the immediate issue is narrower and more concrete: establish whether a covered DeepSeek implementation appears anywhere in the product delivered to an agency, and obtain an authoritative determination before assuming a change of hosting provider makes it permissible.