DeepSeek Is Out of Federal Systems. Is It in Yours? DHS’s Removal Order and the “Built-On” Clause Every Vendor Must Check (AI Trends, 19 October 2026)

What’s trending in AI on 19 October 2026: The US Department of Homeland Security has ordered every civilian federal agency to strip DeepSeek out of its IT systems and stop buying it. That part was expected; agencies from the Navy to NASA had already blocked the Chinese chatbot in early 2025. The surprise is the reach. The order also covers third-party products and services built on DeepSeek’s models, which pulls in any software vendor that quietly routes work through a DeepSeek model, and, through federal contract rules, potentially the thousands of companies that sell to Washington. It lands the same weekend DeepSeek shipped a faster, cheaper open-weight model and a bargain web-search API. Below: what the order says, the clause that matters most, who is actually affected, how a federal contractor’s clock works, and a five-question test to find out whether DeepSeek is hiding somewhere in your stack.

Key takeaways

Remove and don’t buy. DHS ordered civilian federal agencies to remove DeepSeek products and barred future purchases from Hangzhou DeepSeek Artificial Intelligence and its affiliates.
“Built on DeepSeek” counts. The order reaches third-party services developed with DeepSeek’s models, not only the DeepSeek app and API.
It uses a 2018 supply chain law. The order rests on the Federal Acquisition Supply Chain Security Act, after a Federal Acquisition Security Council review, and carries no published removal deadline.
Contractors have their own clock. Federal contract clauses already require suppliers to check for these orders and report affected products within three business days.
Private use is not banned. Businesses outside federal work are not directly covered, but customers and auditors will start asking where your AI models come from.

The hard part of this order is not the chatbot. It is the model layer buried inside other people’s software.

1. What DHS ordered

On Monday, 19 October, DHS filed an order under docket DHS-2026-0628 directing federal civilian executive-branch agencies to remove “covered” DeepSeek products and services from their information systems, and barring them from acquiring any more. AI Weekly reports that a signed copy is on regulations.gov and that the Federal Register notice was due to publish the following day, 20 October. Quartz links it as Federal Register document 2026-21447, titled as an exclusion and removal order for DeepSeek products.

The target is Hangzhou DeepSeek Artificial Intelligence Co., Ltd., and its affiliated entities. DeepSeek is based in Hangzhou and funded by the Chinese hedge fund High-Flyer. The legal basis is the Federal Acquisition Supply Chain Security Act of 2018, usually shortened to FASCSA, and DHS says it acted on a recommendation from the Federal Acquisition Security Council after reviewing the evidence, according to Shattered.io’s summary of the filing.

Three limits matter. First, it covers civilian agencies. None of the coverage we read says it extends to the military, which has run its own DeepSeek blocks since January 2025. Second, no removal deadline has been published; agencies are expected to get implementation guidance from the Office of Management and Budget or their own chief information officers. Third, it is not a ban on private or individual use in the US. You can still download DeepSeek’s app or weights. What changed is that the federal government will no longer run them or pay for anything built on them.

2. Why: three “threat vectors”

The notice speaks of “multiple potential threat vectors,” in The420’s reading, and the coverage consistently lists three:

Data transfer. Sensitive information typed or uploaded by government employees or vendors could end up in the People’s Republic of China.
Intelligence value. Agency use could generate information useful against US persons or government operations, even without any single dramatic leak.
Hidden code. The models or the software around them could contain malware or other vulnerabilities.

Be careful about what the order does not say. The420 notes that the public information does not establish that any government data has actually been transferred, and the notice gives little technical detail. This is a risk-based supply chain decision, the same kind of call governments have made about telecoms equipment and antivirus software, not the conclusion of a breach investigation.

Risks one and two depend on where the model runs. Risk three applies even to self-hosted copies.

3. The clause that matters: “built on DeepSeek”

Banning a chatbot app is easy. Agencies uninstall it and block the domain. The harder part is that the removal requirement covers any third-party service “developed using DeepSeek’s model or its underlying implementations,” in AI Weekly’s quotation of the order. That reaches into the layer most buyers never see: the model a vendor’s product calls behind the scenes.

DeepSeek’s models are popular with software companies for a simple reason: price. We covered the economics in our look at when cheaper Chinese open-weight models make sense. A vendor can call DeepSeek’s API, run its open weights on its own servers, or fine-tune a smaller model on top of them, and the end customer would usually never know. Shattered.io expects US vendors that route some inference through DeepSeek to migrate to other providers if they want to keep federal customers.

There is an open question nobody has answered yet. AI Weekly asks, without resolving it, whether the order extends to fine-tunes and distillations of DeepSeek’s open weights. If a vendor runs a model derived from DeepSeek entirely on US servers, the first two risks DHS cited mostly fall away, because no data leaves the building. The third, hidden flaws in the model itself, does not. Until OMB or agency CIOs publish guidance, the cautious reading for anyone selling to government is that derived models count.

If you sell to government, plan as if the dashed ring is inside the line until guidance says otherwise.

4. Who is actually affected

The headline says “federal agencies,” but the ripple runs wider. Here is how it lands on each group, based on the order and the federal contract rules already on the books.

Who	Directly covered?	What it means now
Civilian federal agencies	Yes	Remove DeepSeek products and built-on services; no new purchases. Deadline still to be set in guidance.
Federal contractors and subcontractors	Yes, through FAR clauses	Check SAM.gov for orders, make a reasonable inquiry into your supply chain, report affected products within three business days.
SaaS and AI vendors selling to government	Yes, if they use DeepSeek models	Swap the model layer or risk losing federal deals; disclose model provenance to customers.
State and local government	Not by this order	Texas, New York and Virginia already restrict DeepSeek on state devices; others may follow the federal lead.
Private businesses and individuals	No	Legal to use, but expect model-origin questions in security reviews and vendor questionnaires.

Sources: Shattered.io, AI Governance and The420.in (19 Oct 2026); Crowell & Moring on FAR 52.204-29 and 52.204-30; Inside Global Tech (Feb 2025).

That last row is where most of our readers sit. AI Governance’s compliance note recommends adding model provenance to vendor due diligence questionnaires and briefing the board on any exposure. Even if you never sell to government, your largest customers might, and they will push the question down to you. It is the same chain reaction we described in the questions to ask your AI vendor after the White House AI Accord.

5. The federal contractor’s clock

Federal suppliers do not need to wait for new rules. Two clauses in the Federal Acquisition Regulation, added by an earlier FAR Council interim rule, already set out what happens when a FASCSA order lands. Law firm Crowell & Moring’s summary of them is the clearest we found:

Representation (FAR 52.204-29). When bidding, you represent that you have made a “reasonable inquiry” and will not supply or use products covered by a FASCSA order, unless you disclose them or get a waiver. The inquiry covers information you already hold; it does not demand a full audit.
Monitoring (FAR 52.204-30). During a contract you must search SAM.gov for “FASCSA order” at least every three months, or more often if your contracting officer says so.
Reporting. If a covered product was or will be delivered or used, report it to the contracting officer within three business days, then update the report with mitigation steps within ten business days.
Flowdown. The rules apply to all subcontracts, including commercial ones. Subcontractors report to the prime, which tells the government.

This machinery has been used once before. In September 2025, the intelligence community issued the first public FASCSA order, against software company Acronis, and Inside Government Contracts noted at the time that relief is limited to review by the D.C. Circuit court of appeals and that agencies can terminate contracts. The DeepSeek order is broader, covering all civilian agencies rather than intelligence systems, and it is the first aimed at an AI model maker.

Three business days is not long if nobody knows which model your vendors call.

6. How we got here

Today’s order is the end of a long ramp. Within days of DeepSeek’s breakout in January 2025, the Navy issued a memo against its models, the Defense Information Systems Agency began blocking it on Pentagon networks, NASA barred it from agency data and devices, and the House told staff not to install it, as Inside Global Tech tracked at the time. Texas, New York and Virginia followed with bans on state devices. In February 2025, Reps. Josh Gottheimer and Darin LaHood introduced the bipartisan No DeepSeek on Government Devices Act, which FedScoop reported, and which has not become law.

What changed in between is that DeepSeek became infrastructure. Its open weights now sit underneath other companies’ products, which is why a device ban no longer does the job. It also fits a wider pattern of governments drawing borders around AI, from China’s draft export controls on model weights and agents last week to the espionage campaigns aimed at AI know-how earlier this month.

Device bans stopped the app. The 2026 order goes after the model wherever it is embedded.

7. Why the temptation is not going away

The timing is awkward for anyone hoping DeepSeek would fade. Over the weekend the company posted DeepSeek-V4.1-Flash-1017 on Hugging Face under the MIT licence, an update aimed at faster agent and coding work with shorter outputs. Its model card shows gains over the previous Flash release, for example 42.3 versus 31.2 on Terminal-Bench 4.0 and 77.0 versus 74.2 on DeepSWE v1.1, while still trailing Anthropic’s Opus 5 on most listed benchmarks. DeepSeek also opened its web search system to developers at $0.75 per 1,000 successful searches; The Neuron’s digest compares that with $5 per 1,000 for Brave’s search API.

Cheap, capable and freely licensed is a hard combination to resist, and that is exactly why the model ends up inside other products. It also shows up in less welcome places: in this lunchtime’s story on the Cairn agents that broke into more than 100 websites, the attacker’s worker was labelled for DeepSeek V4.1 Flash. None of that makes the model itself malicious; open weights can be used by anyone. But it is a reminder that the question “which model is this?” is now a security question, not just a procurement one. We made a similar point about China’s GLM-5.3 and open-weight hacking skills.

8. Is DeepSeek in your stack? Five questions, in order

Recent guides here have been role-split checklists and week-by-week plans. This one is a decision path: work down the questions, and stop at the first “yes” to act on it before moving on. It is general guidance, not legal advice; federal contractors should involve counsel and their contracting officer.

Does anyone on staff use the DeepSeek app or website for work? If yes: block it on managed devices and networks, and point people to an approved alternative. Unapproved chatbots are the classic shadow AI problem, and your data policy should name which tools are allowed.
Does any of your own code call DeepSeek’s API or run its weights? If yes: search code and infrastructure for DeepSeek API endpoints, keys and model names, list every workload, and decide model by model whether to switch providers, especially for anything touching customer data.
Do any of your AI vendors use DeepSeek underneath? If you don’t know: ask in writing which base models power each AI feature, where inference runs, and whether any model is a fine-tune or distillation of a DeepSeek model. Add those questions to your standard vendor review.
Do you hold, or bid for, federal contracts or subcontracts? If yes: search SAM.gov for the DeepSeek FASCSA order, run a reasonable inquiry using answers from steps 1 to 3, and be ready to report affected products to the contracting officer within three business days, with a mitigation update within ten. Flow the requirement to your own subcontractors.
Do your customers sell to government, or work in a regulated sector? If yes: prepare a one-paragraph model provenance statement now: which models you use, where they run and who hosts them. You will be asked, and a ready answer wins deals.

If every answer is no, you still have one task: record the result and the date, and repeat the check each quarter. That habit costs little, and it matches the three-month rhythm federal contractors already follow. For the wider picture of who owns AI risk inside a company, see our guide to AI agent security.

9. The bottom line

The DeepSeek order is less about one Chinese chatbot than about a new expectation: that organisations know which AI model sits inside each product they buy. For federal agencies and their suppliers that expectation is now enforceable, with a three-business-day reporting clock already written into contract law. For everyone else it is arriving through customer questionnaires and security reviews. The unresolved question, whether fine-tuned or self-hosted derivatives count, will be settled by guidance in the weeks ahead. You do not need to wait for it to find out what your own stack runs on.

Frequently asked questions

Did the US ban DeepSeek?

Not for the public. On 19 October 2026, DHS ordered federal civilian agencies to remove DeepSeek products and third-party services built on its models from their IT systems and barred them from buying more. Private companies and individuals in the US are not directly covered.

What law is the DeepSeek order based on?

The Federal Acquisition Supply Chain Security Act of 2018 (FASCSA). DHS acted after a recommendation from the Federal Acquisition Security Council. The order is filed under docket DHS-2026-0628, with a Federal Register notice due on 20 October 2026.

Does the order cover software that only uses DeepSeek behind the scenes?

Yes, as reported. The removal requirement extends to third-party products and services developed using DeepSeek’s models or their underlying implementations. Whether that includes fine-tunes or distillations of DeepSeek’s open weights running on US servers has not yet been clarified.

What do federal contractors have to do?

Under FAR 52.204-30, contractors must check SAM.gov for FASCSA orders at least every three months, make a reasonable inquiry into their supply chain, report covered products to the contracting officer within three business days, and update that report with mitigation steps within ten business days. The rules flow down to subcontracts.

Is there a deadline for agencies to remove DeepSeek?

No removal deadline has been published. Implementation guidance is expected from the Office of Management and Budget or agency chief information officers.

Should a private business stop using DeepSeek?

That is a risk decision, not a legal requirement, unless you do federal work. Weigh where your data goes, whether customers or regulators will ask about model origin, and whether you sell into government supply chains. At a minimum, know where DeepSeek models are used in your systems and your vendors’ products.