DHS Purges DeepSeek — and Everything Built on It
Washington has finally stopped warning about DeepSeek and started ripping it out. On Monday, the Department of Homeland Security ordered every civilian federal agency to purge DeepSeek products from its IT systems — and went further than any American restriction before it: the order follows the Chinese AI model into the supply chain, sweeping up any third-party service built using it.
This is not a ban on an app. It is a ban on a model, wherever it hides — and that makes it the deepest, and by far the hardest to enforce, AI crackdown the US government has ever attempted. If TikTok was the dress rehearsal, this is opening night of a much uglier play.
Here is what landed on the Federal Register website Monday morning, per the notice and same-day reporting: federal civilian agencies must remove any DeepSeek products from their systems; they are banned from acquiring DeepSeek products going forward; and — the part nobody in Washington had done before — the acquisition ban extends to third-party services developed using DeepSeek. Bloomberg Law separately reported that the order bars the Chinese firm from participating in federal government contracts.
What the order actually does
The mechanics are blunt. A notice posted Monday by DHS on the Federal Register website, citing US cybersecurity laws, does three things at once: it orders the removal of DeepSeek products and third-party services using the Chinese AI model from civilian agency IT systems; it bans future acquisition of DeepSeek products and third-party services developed using DeepSeek by those agencies; and, per Bloomberg Law, it shuts the Chinese company out of federal government contracting.
The language DHS chose is unusually direct for a procurement notice. The department said DeepSeek models pose "multiple threat vectors" and raised concerns that use of the product "may result in the transfer of sensitive data to the PRC if used by U.S. government employees or vendors." Note the last word: vendors. This is not aimed only at bureaucrats fiddling with a chatbot on a work laptop. It is aimed at the contractors, subcontractors, and software suppliers whose tools sit on federal networks — and whose tools might quietly be calling a DeepSeek model under the hood.
One caveat worth keeping straight: this is a civilian-agency order. DHS directives of this kind cover the federal civilian executive branch; the Pentagon and the intelligence community already live under their own, harsher DeepSeek rules (more on those below). That split matters, because it tells you what this order is really about — protecting the vast, boring, data-rich machinery of civilian government: benefits databases, tax systems, immigration records, health data.
The twist: Washington is chasing the weights, not the app
Banning an app is easy. You delist it, you block the domain, you write a memo. Banning a model — an open-weight model that anyone can download, self-host, rebrand, and pipe through an API with the DeepSeek name nowhere in sight — is a different animal entirely. And the administration has apparently decided to hunt it anyway.
That is what the "third-party services developed using DeepSeek" clause is doing. A vendor selling an analytics tool to, say, a civilian agency could be running its summarisation layer on a hosted DeepSeek model. The agency's procurement officers would never see the word "DeepSeek" on the invoice. Under Monday's order, that service is now prohibited — which means federal IT shops suddenly have to ask their vendors a question Washington has never forced them to ask: whose model weights are actually running in your product?
We read the notice's reported language carefully on this point, because it is the entire ballgame. The administration is treating the model weights themselves as a potential exfiltration vector — the idea being that even a locally hosted copy of a model developed under Chinese jurisdiction can be shaped, steered, or subtly compromised in ways that leak data back to Beijing. Whether you find that paranoid or prudent depends on how much you trust the House Select Committee's findings (below), but as a doctrine it is new: the US government is no longer just sanctioning companies. It is sanctioning matrices.
That doctrine has been building for months. When we covered Rep. Ro Khanna grilling AI CEOs on whether China is coming for their model weights, the anxiety in the hearing room was about exactly this: the weights are the crown jewels, and nobody is entirely sure whose hands they pass through on the way to a federal contract.
Eighteen months of siege: how DeepSeek became target number one
Monday's order did not come out of nowhere. It is the end — or at least the latest chapter — of an eighteen-month campaign that began the moment DeepSeek's R1 model rattled global markets in January 2025, triggering a major equity sell-off as investors realised a Chinese lab had matched Silicon Valley's frontier models at a fraction of the cost.
The first bans were the small, fast ones. In early 2025, the House of Representatives' Chief Administrative Officer barred congressional staff from installing or using DeepSeek on official devices, with a notice warning that the app was "under review" and "currently unauthorized for official House use" — and that "threat actors are already exploiting DeepSeek to deliver malicious software and infect devices." The US Navy issued its own directive to personnel. The Commerce Department banned it on government devices. New York, Virginia, and Texas followed at the state level.
Then came the legislation. In February 2025, Reps. Josh Gottheimer (D-N.J.) and Darin LaHood (R-Ill.) introduced the "No DeepSeek on Government Devices Act." Gottheimer did not mince words: "The Chinese Communist Party has made it clear that it will exploit any tool at its disposal to undermine our national security, spread harmful disinformation, and collect data on Americans," he said in a statement. The bill had exceptions for national-security research — an early sign that even the hawks understood you cannot study an adversary's AI without touching it.
The heaviest artillery arrived in April 2025, when the House Select Committee on China published its report, "DeepSeek Unmasked: Exposing the CCP's Latest Tool for Spying, Stealing, and Subverting US Export Control Restrictions." Its findings read like an indictment: DeepSeek, the committee said, "acquires extensive personal data on the Americans who use the chatbot, including chat history, device details, and even the way a person types" — and then, "by its own admission, funnels the data directly back to China." The chatbot, it found, alters or suppresses responses to topics deemed politically sensitive by the CCP in 85% of cases. It was "highly likely," the committee concluded, that DeepSeek had distilled US models to build its own, a charge American AI companies told investigators they had "high confidence" in. And the models "appear to be powered by" tens of thousands of Nvidia chips that were supposed to be restricted from export to China.
"DeepSeek represents a profound threat to our nation's security," the report said. "Although it presents itself as just another AI chatbot … closer inspection reveals that the app siphons data back to the People's Republic of China (PRC), creates security vulnerabilities for its users, and relies on a model that covertly censors and manipulates information pursuant to Chinese law."
Here is the human detail that stuck with me from that spring: a day after Washington announced it would require a licence to export H20 chips to China, Nvidia founder and CEO Jensen Huang was in Beijing — a surprise visit, meeting senior officials at the invitation of the state-backed China Council for the Promotion of International Trade, according to state-run CCTV. The chip king paying court in the capital while Congress drafted his company's interrogation. You could not have staged the tension better.
The drumbeat never stopped. In March of this year, the Wall Street Journal reported that the Trump administration was likely to ban DeepSeek from government devices outright — and was also weighing pulling the chatbot from app stores and limiting how US cloud providers could offer its models to customers. Those wider discussions, the Journal cautioned, were "still at an early stage." On March 6, 21 state attorneys general wrote to congressional leaders urging them to pass a federal ban, calling China "a clear and present danger" and DeepSeek "another tool for Chinese spies to attack America's national security." Montana Attorney General Austin Knudsen, who drafted the letter, put it more colorfully: "China is trying to steal America's secrets. America must fight back. Congress should shut down China's latest Trojan horse by passing the No DeepSeek on Government Devices Act."
The military, meanwhile, had already moved. Section 1532 of the FY2026 National Defense Authorization Act requires the Secretary of Defense to exclude and remove AI developed by DeepSeek or its parent company, the quant hedge fund High-Flyer, from Pentagon systems within 30 days of enactment — and forces defense contractors to swear off "covered AI" in contract performance, with only narrow waivers for research and counterintelligence work. That provision turned model-lineage tracing from a best practice into a compliance obligation running down every tier of the defense supply chain.
Monday's DHS order is the civilian half of that same wall finally going up.
Why DHS, and why now
The timing is worth a raised eyebrow. This is the same administration that has been threatening devastating retaliation over EU curbs on Starlink — a White House perfectly happy to escalate tech fights on multiple fronts at once. And it comes as the US and China have been inching toward something resembling a trade thaw, with Chinese state-owned firms buying up American soybeans in the wake of the recent Trump-Xi summit. The message seems to be that commerce is negotiable and AI infrastructure is not.
As for the instrument: DHS is the natural home for this. Civilian agencies are where the data lives — and where the procurement mess lives. A civilian agency's IT stack is a sedimentary layer cake of legacy systems, cloud contracts, and vendor tools, assembled over decades. Telling a hundred-odd agencies to find and remove not just an app but any service developed using a particular foreign model is, to put it mildly, an ambitious homework assignment.
We checked the Federal Register's public-inspection pages on Monday afternoon; the notice's document number had not yet surfaced in the retrievable public record at press time, so this account is built on the notice language as carried by Reuters and Bloomberg Law rather than the full text. When the complete document posts, the details that matter will be the compliance timelines and the exact definition of "developed using DeepSeek" — the two phrases on which the entire enforcement effort will live or die.
The enforcement problem: you cannot firewall a download
And now the part where I stop narrating and say what I actually think: this is simultaneously the strongest move Washington has made against DeepSeek and the one most likely to be colossally difficult to enforce — and the administration should be honest about that instead of pretending a Federal Register notice can contain an open-weight model.
The record here is not encouraging. Washington has been trying to purge "covered" Chinese technology from federal supply chains for years, and the results have been shoddy. Section 889 of the 2019 NDAA banned Huawei and ZTE gear from federal contracting, and agencies spent years playing whack-a-mole with relabeled equipment buried three tiers deep in vendor stacks. Now multiply that problem by the nature of AI models: a weight file is not a router. It can be renamed, fine-tuned beyond recognition, distilled into a new model, served through a friendly API, and embedded in a product whose vendor has never heard of Hangzhou. Asking every civilian agency to certify that no service it buys was "developed using DeepSeek" is asking for an audit capability the federal government has never demonstrated — against a target designed, by its open-weight nature, to be unfindable.
That does not make Monday's order wrong. The Select Committee's findings on data funnelling and censorship are serious, the "multiple threat vectors" framing is defensible, and there is a coherent argument that the government should not be paying vendors who pay — directly or indirectly — a company whose data pipeline ends in Beijing. But a ban is only as strong as its enforcement, and enforcement here requires something Washington has botched repeatedly: the ability to see inside its own supply chain. Until DHS publishes the compliance machinery — the timelines, the vendor attestation regime, the penalties — this reads less like a wall and more like a very sternly worded fence. A fence with excellent intentions and no gate guard.
There is also the irony the administration would rather not dwell on: DeepSeek, the national-security menace, is simultaneously a commercial juggernaut raising serious money — the same company we recently reported is putting $400 million into a stealth robotics startup. Washington can bar it from federal contracts; it cannot bar it from capital markets, from open-source mirrors, or from the millions of developers who will keep downloading its weights regardless of what the Federal Register says.
What to watch next
Three things. First, the full notice text: the compliance deadline and the precise scope of "developed using DeepSeek" will determine whether this is a surgical instrument or a blunt one. Second, the two shoes the Journal reported back in March — app-store removal and cloud-provider limits. Monday's order is the government-device shoe dropping; if the administration now moves on app stores or US clouds hosting DeepSeek models, that is the moment this goes from a federal procurement story to a consumer-market story. Third, Beijing's response: China has called past restrictions the "politicisation of trade and tech issues," and a formal protest — or a retaliatory move against a US firm — would not surprise anyone.
One more thing to watch, closer to home: whether any agency actually finds DeepSeek-derived services in its stack and says so publicly. The first disclosure will tell us more about the real penetration of these models into government IT than any committee report has.
Sources: Reuters (reporting by David Shepardson, Oct 19, 2026); Bloomberg Law (Oct 19, 2026); House Select Committee on China, "DeepSeek Unmasked" (Apr 2025, via Mondaq/Business Standard/PTI summaries); Wall Street Journal reporting via Economic Times (Mar 2026); letter of 21 state attorneys general via Business Insurance (Mar 6, 2026); EveryCrsReport IF13197 and LexBlog/Petronella on FY2026 NDAA §1532; TMTPost/Axios on the House CAO directive; Federal Register public-inspection pages (checked Oct 19, 2026).
Daniel Rao is a senior correspondent covering AI policy and the US-China technology contest. He has reported on export controls, model security, and federal AI procurement for TechPulse USA.
Update, Oct. 19: the paper trail and the next shoe to drop
Hours after the post went live, the paper trail got sharper. The Federal Register public-inspection filing for the order is document 2026-21447, under docket DHS-2026-0628, filed at 8:45 a.m. Monday and scheduled for official publication on Tuesday, October 20, per the filing's inspection record. In a July memo made public Monday, DHS also told federal agencies to monitor traffic flowing in and out of government websites to detect DeepSeek-related risks, per Reuters' David Shepardson. And the crackdown is not stopping at DHS: Reuters reports the FCC will vote on October 29 on barring all Chinese labs from testing electronic devices for US use — last year, 82% of US electronics were tested in China — after the agency banned imports of new Chinese-made drones and consumer routers in December. For readers tracking the same Washington–Beijing AI fault line: Reuters separately found Alibaba's Qwen powering search on a federal agency's site earlier this month, and Anthropic has accused Chinese startup Zhipu of copying Claude.
DHS DeepSeek ban FAQ: what the order covers — and what it doesn't
Does the DHS order cover the military and intelligence agencies?
No. The order applies to federal civilian executive-branch agencies, except national-security systems and secure compartmented information systems, which fall under the Pentagon and the Director of National Intelligence.
Does the ban reach apps built on DeepSeek's open-weight models?
That is the order's widest clause — it covers "third-party products and services developed using DeepSeek implementations." The order does not list which products it means, and does not say how agencies should treat a DeepSeek model running on a US company's servers.
Is this the first US ban on DeepSeek?
Hardly. The Department of Commerce and the Navy had already banned DeepSeek on government devices, and multiple states — including New York, Tennessee and Virginia — took similar steps earlier.
What did DHS say the threat actually is?
"Multiple threat vectors," including possible transfer of sensitive data to China if government employees or vendors use the models, the chance of producing "actionable intelligence on US persons," and the warning that the models themselves "may contain malware or other security vulnerabilities."
Comments
Post a Comment