DHS Orders Civilian Agencies to Purge DeepSeek From Government Systems

The Department of Homeland Security has ordered every federal civilian agency to stop buying products from the Chinese artificial intelligence company DeepSeek and to strip existing DeepSeek technology out of government computer systems, according to a notice published Tuesday in the Federal Register. The directive, first reported by TechCrunch, is among the broadest actions any part of the US government has taken against a single foreign AI company, reaching past DeepSeek's own chatbot and API to cover any third-party product built on its technology.

Formally called Exclusion and Removal Order 2026-01, it was issued by Homeland Security Secretary Markwayne Mullin under the Federal Acquisition Supply Chain Security Act of 2018, a law that lets DHS bar information-technology products it deems a risk from federal civilian systems. A copy of the signed order has been posted to the public docket alongside the Federal Register filing.

What the order requires

The order gives agencies two obligations. First, they must immediately exclude DeepSeek, its corporate parent Hangzhou DeepSeek Artificial Intelligence, and any affiliated entity from future procurements, including approved-product lists, equipment contracts and online marketplaces used to buy hardware and software. Second, they must remove any DeepSeek product already in use, along with any third-party product or service "developed using DeepSeek implementations," from their information-technology systems, or request a formal exception.

Agencies seeking an exception have 60 days from the order's issuance to submit a written request to DHS's Cybersecurity and Infrastructure Security Agency, laying out a "compelling justification," such as a mission-critical function that cannot otherwise be performed. CISA will review and approve or deny each request. In the meantime, DHS is telling agencies that still have DeepSeek on their networks to tighten monitoring of outbound traffic, strengthen logging and watch for signs of data exfiltration while removal is underway.

The authority behind the order traces back to the Federal Acquisition Supply Chain Security Act of 2018, which created an interagency body called the Federal Acquisition Security Council to flag risky IT products and recommend they be barred from federal procurement and networks. Under that law, the Homeland Security Secretary can act on the council's recommendations for civilian agencies, while the Pentagon and the intelligence community handle exclusion orders for their own systems. DHS says its order followed a council recommendation and a review of what it calls "affiliated evidence" on DeepSeek's data practices.

A delayed disclosure

The order itself carries an "original release date" of July 24, 2026, roughly three months before DHS made it public. That gap, which Engadget noted the department has not explained, means agencies were technically bound by the restrictions well before this week's Federal Register notice. The order excludes national-security systems and sensitive compartmented information systems, which fall under the Secretary of Defense and the Director of National Intelligence rather than DHS, and it requires the Homeland Security Secretary to review the order within one year and annually after that, as the 2018 supply-chain law requires.

In the order, DHS lays out its reasoning in unusually specific terms. It says DeepSeek's servers are located in China, where national-security law gives the government "extensive authority" to reach into private companies' infrastructure, and that the company's account-creation process collects personal data and can track and fingerprint users. Because DeepSeek's chatbot works by taking in whatever a user types and returning a response, the department argues that any sensitive information submitted by a federal employee or contractor could end up accessible to Chinese authorities. "DeepSeek's models present multiple threat vectors," the order states, warning that the resulting data, combined with other available information, "could produce actionable intelligence on U.S. persons and/or government programs." It also cites outside research finding that DeepSeek's iOS app contained multiple security vulnerabilities.

How Washington got here

DeepSeek has been under scrutiny in Washington since its R1 reasoning model, released in January 2025, rattled markets and triggered a sharp sell-off in Nvidia shares by showing that a competitive model could reportedly be trained far more cheaply than its US rivals. Individual agencies moved quickly afterward: the Pentagon's defense information network blocked access within days, NASA restricted employee use, and the Commerce Department banned the app on government devices. Those were agency-level decisions, though, made under each department's own authority.

Congress tried, and failed, to make a government-wide ban into law. Bills such as the No DeepSeek on Government Devices Act stalled in both chambers. Lawmakers did succeed in narrower terms: a defense policy law enacted in December 2025 barred DeepSeek from systems used by the intelligence community and national-security agencies. This week's DHS order extends that restriction to the much larger population of civilian agencies that the earlier law did not reach, using existing supply-chain-security authority rather than new legislation.

The order also lands amid a broader dispute over whether Chinese AI labs are improperly studying American models. A September report accused DeepSeek and five other Chinese companies, including Alibaba and Moonshot AI, of large-scale "distillation," a technique that trains a new model on another model's outputs. Both OpenAI and Anthropic have separately said they detected unauthorized attempts by DeepSeek and other Chinese labs to harvest their systems' outputs this way. China, for its part, is reportedly weighing its own export controls on AI-related technologies, a move Bloomberg reported on the week before the DHS order became public.

Who is affected

The order applies only to federal civilian executive-branch agencies, not the Defense Department or the intelligence community, which already operate under the December 2025 restrictions and their own security rules. But "civilian" covers the bulk of the federal government, from the Treasury and HUD to the EPA and the Social Security Administration, along with the contractors and software vendors that sell those agencies IT products and services.

The clause requiring removal of "third-party products and services developed using DeepSeek implementations" is the part industry lawyers are expected to scrutinize most closely. DeepSeek releases its models as open weights, meaning any developer can download, modify and repackage them, and a wide range of commercial AI tools and open-source projects have incorporated DeepSeek's models or distilled versions of them without necessarily advertising that origin. Vendors supplying the federal government may now need to trace and attest to whether their products rely on DeepSeek technology anywhere in their stack, a process cybersecurity commentators have already flagged as difficult to verify in practice.

Beijing's response

China's government pushed back the same day. Asked about the order at a regular briefing, Foreign Ministry spokesperson Mao Ning said Beijing opposes "overstretching the concept of national security and politicizing technology and trade issues."

"We will resolutely safeguard the legitimate rights and interests of Chinese enterprises," Mao said.

DeepSeek had not issued its own public statement addressing the order as of Tuesday morning.

What happens next

With the formal notice now filed, the 60-day clock for agencies to request exceptions is running, and CISA is expected to begin fielding and ruling on those requests through its FASCSA order-implementation team. DHS has not published a public list of which agencies or systems still contain DeepSeek technology, so the practical scope of the removal effort is likely to become clearer only as individual agencies report their progress or seek exceptions.

The order is unlikely to be the government's last word on Chinese open-weight AI. CyberScoop and other outlets covering the order noted that other Chinese models drawing similar distillation and security concerns, including Alibaba's Qwen and Moonshot AI's Kimi, were not named in this action, leaving open the question of whether DHS or Congress extends comparable restrictions to them. For now, the immediate fight is likely to center on the murkier question the order raises for government contractors: how to prove a product has no DeepSeek inside it at all.