The US Barred DeepSeek From Federal Agencies — Reportedly Including Products Built on It. Do You Know What Your Product Is Built On?

It's in the layersHover to explore

Most AI products do not use one model. They use a hosted API for the main feature, an open-weight model for something cheap, an embedding model nobody remembers choosing, and a coding assistant in the team's editor. Some of those models were fine-tuned or distilled from others. A US order on DeepSeek, made public this week, is the first time many founders will be asked to account for all of it.

This article is based on reporting by CyberScoop, Security Magazine, PYMNTS and Tech Insider, and on the Department of Homeland Security's announcement as they describe it. I have not read the Federal Register notice itself, and the outlets differ on some details, noted below. The reading from section 3 onward is mine, and it is not legal advice.

1. What the order does

On 19 October 2026, DHS announced an order under the Federal Acquisition Supply Chain Security Act of 2018 covering Hangzhou DeepSeek Artificial Intelligence and its affiliates. As reported, it requires federal civilian agencies to:

• remove DeepSeek products from their information-technology systems; and
• stop acquiring them in future.

Exemptions require 60 days' advance notice and a "compelling justification", according to CyberScoop. The Department of Defense is reported to be outside the order, which applies to civilian agencies. The notice was due to be published in the Federal Register this week.

DHS Secretary Mullin cited "multiple threat vectors", including "data in servers based in China, where national security laws give the Chinese government wide latitude to request and obtain data", the collection of personal information and digital fingerprints, and data passing to the Chinese government through the "prompt and response" process. Reporting also lists the risk of malware or vulnerabilities introduced through the models.

Private companies are not banned from using DeepSeek by this order.

2. The clause that matters to vendors

Several outlets, citing DHS reporting by CNBC and Politico, describe the order as also barring agencies from buying "third-party products and services developed using DeepSeek's technology or AI models." That would close the gap that earlier, narrower restrictions left open: a vendor's chatbot, coding assistant or summariser running on a DeepSeek model, even several layers removed from DeepSeek itself.

Not everyone reports it the same way. CyberScoop's account describes the ban on DeepSeek's own products and says the treatment of third-party products, self-hosted open weights, fine-tunes and distillations is not explicitly addressed. Until the notice text is read closely, the honest position is:

• Reported: third-party products built on DeepSeek models are covered.
• Unclear: whether that includes open weights you host yourself, models fine-tuned from DeepSeek, or models distilled from DeepSeek outputs.

3. Why this reaches startups that never sell to government

You may not sell to federal agencies. But three groups of buyers often follow federal rules:

• Government contractors, who carry these requirements into their own vendor reviews.
• Regulated enterprises — banks, healthcare, critical infrastructure — whose security teams treat federal exclusions as a baseline.
• State and local governments, several of which had already restricted DeepSeek before this order.

For a startup selling into any of them, the question on the next security questionnaire is easy to predict: does your product use, or depend on, any DeepSeek model? The hard part is answering it truthfully.

4. Why the answer is harder than it looks

A "no" has to hold at every layer:

• The APIs you call. Some inference providers and model routers serve DeepSeek models alongside others, and a router can switch models without a code change.
• The open-weight models you host. Several widely used models are fine-tunes or distillations of DeepSeek releases, and the lineage is sometimes only in a model card.
• Your team's tools. Coding assistants and agents in your engineers' editors may be configured to use DeepSeek, even if your product never does.
• Your vendors. A transcription, search or support tool you embed may run on a model you cannot see.

Our guides to DeepSeek V4-Flash and Qwen3.8-Max covered why open-weight models are attractive on price and quality. This order does not change that calculation for most private uses. It changes what you need to be able to prove to some customers.

5. What to do this month

If a customer needs DeepSeek-free, you will only be able to promise it if you can see your whole stack. If they do not, you will still be able to tell them exactly what they are getting — which is usually what a security team wants.

6. What I would not claim

That startups are banned from using DeepSeek. The order applies to federal civilian agencies and what they buy. Private use is not prohibited by it.

That self-hosted, fine-tuned or distilled models are definitely covered. The reported wording is broad, but the outlets disagree, and the notice text should be read before deciding.

That the dates are settled. CyberScoop reports the order was issued in June and has been in force for months; other outlets give July; publication in the Federal Register was reported for 20 or 21 October.

That DeepSeek's models are unsafe to run. The order is about supply-chain and data-access risk as assessed by the government. It is not a technical finding about any particular model's weights.

The honest summary

DHS has ordered federal civilian agencies to remove DeepSeek products from their systems and stop buying them, under the Federal Acquisition Supply Chain Security Act, citing the risk of data reaching the Chinese government. Several outlets report that the order also covers third-party products built on DeepSeek's models; how far that reaches into self-hosted, fine-tuned and distilled models is not yet clear. Private companies are not banned.

The lasting effect for startups is not about DeepSeek alone. Buyers who follow federal rules will start asking which models are inside your product and where they came from, and the usual answer — "whatever the API is using" — will not pass.

This month, write down every model in your product and your team's tools, including what each was built from. Pin them in code, log which one served each request, and make sure you can replace any one of them in days.

Sources: CyberScoop — the Federal Acquisition Supply Chain Security Act basis, the requirements to remove and stop acquiring DeepSeek products, the civilian-only scope, the 60-day exemption notice, DHS Secretary Mullin's description of the threat vectors, the reported June issue date, and the statement that third-party, self-hosted, fine-tuned and distilled models are not explicitly addressed. Tech Insider — the reported wording on "third-party products and services developed using DeepSeek's technology or AI models", attributed to CNBC and Politico, the July effective date it gives, and the exclusion of the Department of Defense. Security Magazine and PYMNTS — the 19 October announcement and coverage of the order's scope. The reading in sections 3 to 6 is mine and is not legal advice.

Related Posts

DeepSeek's Cheap Model Just Beat Its Own Flagship: What V4-Flash 0731 Means for Your StackAugust 4, 2026

Alibaba Open-Sourced a 2.4-Trillion-Parameter Model and Priced It Like a Closed OneAugust 6, 2026

The FTC Asked for Incidents and Marketing in the Same LetterOctober 16, 2026