U.S. orders federal agencies to purge DeepSeek over security concerns

The U.S. government has ordered federal civilian agencies to stop buying DeepSeek products and services and to remove the Chinese AI startup’s technology from government systems, citing national security risks tied to how the company operates in China. The directive broadens Washington’s effort to keep DeepSeek out of sensitive networks and extends restrictions beyond the company’s own tools to third-party products built on its technology.

The move marks one of the strongest U.S. responses yet to the rapid spread of DeepSeek, whose low-cost open-weight models jolted the AI market earlier this year and intensified the geopolitical contest between Washington and Beijing over advanced artificial intelligence.

What the order does and why it matters

The new instruction tells federal civilian agencies to stop purchasing DeepSeek offerings and to scrub the startup’s technology from their IT environments, according to the Department of Homeland Security. It also applies to products and services made by other vendors if they rely on DeepSeek technology.

That broader scope matters because it aims to close off indirect use, not just direct procurement. In practice, the order could affect a wider ecosystem of software tools, integrations, and services that may quietly depend on DeepSeek models or infrastructure.

Agencies may still seek exceptions, but the default position is clear: DeepSeek is now treated as a security risk inside the civilian federal government.

Why the U.S. is acting now

The U.S. government’s concern centers on data handling and the possibility that information entered by American officials could be reachable under Chinese law. DHS said DeepSeek’s operations in China raise the risk that sensitive government data could be accessed by Chinese authorities.

In national security circles, the issue is not only whether a foreign company stores data securely today, but whether legal and regulatory obligations in its home country could force access later. That concern has become a key fault line in U.S.-China tech policy, especially for companies handling large volumes of text, code, and internal government material.

DHS warned that data submitted by U.S. government users could potentially be available to Chinese authorities because of the company’s ties to China and the legal environment in which it operates.

The order reflects a broader Washington strategy: block foreign AI systems from embedded use in government before they become deeply integrated into official workflows.

How the DeepSeek ban expands earlier restrictions

This is not the first U.S. action against DeepSeek. A law enacted in December 2025 already limited the use of the company’s technology in intelligence and national security systems. The latest step pushes the restriction into the much larger world of federal civilian agencies, and it covers existing systems as well as future purchases.

That distinction is important. Banning new contracts is one thing; ordering removal from live systems is more disruptive. It suggests federal officials believe the risk is significant enough to justify the cost and complexity of migration.

How DeepSeek became a geopolitical flashpoint

DeepSeek first drew global attention when its R1 reasoning model stunned the AI industry in January 2025. The model was seen as evidence that Chinese labs could compete aggressively with U.S. leaders while using a fraction of the cost and compute that had previously been assumed necessary for frontier AI development.

The reaction in financial markets was immediate. Nvidia shares dropped sharply after the release, underscoring how quickly a breakthrough from a relatively little-known Chinese lab could unsettle expectations about the economics of AI infrastructure.

Since then, DeepSeek has become part of a much larger strategic rivalry. To U.S. policymakers, the company is not just another AI startup; it is a test case for whether America can keep adversarial technologies out of sensitive networks without slowing down innovation at home.

Why open-weight models raise the stakes

DeepSeek’s systems are especially consequential because they are low-cost and open-weight, which makes them easier to adapt, redistribute, and embed into other products. That accessibility can accelerate adoption, but it can also make it harder for governments to track where the technology is being used.

For security teams, open-weight distribution creates a long tail of potential exposure. A model can be downloaded once and then re-used inside a wide variety of applications, from internal chatbots to document analysis tools. That makes enforcement more complicated than a simple app-store ban.

What concerns have U.S. officials raised about data and model behavior?

U.S. officials have focused on two broad issues: data exposure and model extraction. The first is the risk that information entered into DeepSeek systems could be collected, retained, or made available under Chinese legal obligations. The second is the allegation that DeepSeek and other Chinese AI firms have sought to learn from American AI systems through model distillation.

Model distillation refers to a process in which outputs from one model are used to train another, allowing a newer system to imitate or approximate a competitor’s behavior. In highly competitive AI markets, that technique can amount to a shortcut around years of research and enormous training costs.

The U.S. accused DeepSeek and five other Chinese AI companies, including Alibaba and Moonshot AI, of using large-scale distillation to extract information from American AI systems, according to a September report. DeepSeek has not publicly addressed all of those allegations in the material provided.

What OpenAI and Anthropic have said

Separate from government action, leading U.S. AI companies have also voiced concerns. OpenAI told lawmakers earlier this year that it had observed suspected attempts by DeepSeek to distill its models. Anthropic said it had identified unauthorized attempts by DeepSeek and other Chinese AI labs to collect outputs from Claude.

OpenAI and Anthropic have both framed the issue as a misuse of model outputs, describing it as an attempt to learn from or replicate proprietary systems without authorization.

Those complaints matter because they show that the debate is not limited to national security agencies. Commercial AI leaders now see model extraction as a business and intellectual property threat as well as a geopolitical one.

How China responded to the U.S. decision

China rejected the U.S. move and accused Washington of politicizing technology and trade. Foreign Ministry spokesperson Mao Ning said China opposes the “overstretching” of national security arguments and vowed to defend the rights and interests of Chinese companies.

The response fits a familiar pattern in the tech confrontation between the two countries: the United States frames the issue as protection of infrastructure and data, while China portrays restrictions as unfair attempts to block its companies from global markets.

That disagreement is likely to sharpen as more AI products are embedded in government and enterprise systems. When tools become part of everyday administrative work, security disputes can quickly turn into trade and diplomacy disputes.

What does this mean for federal agencies?

Federal civilian agencies will likely need to audit software inventories, identify any direct or indirect DeepSeek dependencies, and replace them with approved alternatives. That includes not only standalone AI services, but also any vendor products that may have quietly integrated DeepSeek models behind the scenes.

For many agencies, the hardest part may be detection. AI models are often bundled inside broader platforms, and procurement records do not always reveal the origin of the underlying model. The new order could therefore force a wave of compliance reviews across contractors, IT teams, and procurement offices.

Likely operational steps

• Review all contracts and software subscriptions for DeepSeek use.
• Inventory internal systems that connect to DeepSeek APIs or models.
• Identify third-party tools that rely on DeepSeek under the hood.
• Submit exception requests where mission-critical use is claimed.
• Replace or retire affected systems where exceptions are denied.

Because the order includes exceptions, agencies with specialized needs may still argue for limited use. But the burden now shifts to officials to justify why a restricted AI tool should remain in place.

Why this is part of a larger U.S.-China AI rivalry

The DeepSeek decision should be read as one move in a much larger contest over the future of artificial intelligence. The U.S. and China are competing not only on model quality, but on compute supply chains, chip access, model distribution, safety rules, and government procurement.

DeepSeek’s rise has complicated assumptions that U.S. firms would maintain a stable lead simply by spending more on training. If a Chinese company can offer capable models at lower cost, the strategic challenge becomes bigger than market share. It becomes a question of whether the U.S. can preserve trust in the systems used by its own government.

That is why the debate has moved beyond pure performance metrics. In Washington, the central issue is increasingly whether an AI system can be trusted with public-sector data, not just whether it can write better code or answer questions more cheaply.

Could China tighten its own controls too?

Yes. China is reportedly weighing export restrictions on 23 AI-related technologies, including chip designs and certain advanced models, according to Bloomberg. If enacted, those controls would add another layer to an already fragmented global AI landscape.

Such restrictions would not be surprising. Governments on both sides of the Pacific have increasingly treated AI as strategic infrastructure, not just commercial software. That means access can be limited through export controls, procurement bans, data rules, or licensing requirements.

The result is a world in which the same AI sector is being pulled in opposite directions: toward global distribution by engineers and startups, and toward national containment by governments.

Timeline of the DeepSeek crackdown

The following timeline shows how quickly DeepSeek moved from a disruptive newcomer to a target of government restrictions:

What happens next?

The immediate next step is enforcement. Agencies will have to determine where DeepSeek is embedded, whether exceptions are justified, and how quickly they can migrate away from affected systems.

Beyond that, the order may influence how U.S. companies, contractors, and software vendors talk about their AI supply chains. If federal buyers begin asking where models come from and how they are trained, the pressure could ripple into the commercial market as well.

DeepSeek’s case also shows that the global AI race is no longer just about who builds the best model. It is now about who gets to trust which systems, where the data flows, and which governments can draw the security lines first.

For Washington, the message is unmistakable: even if a foreign AI model is affordable and competitive, that alone will not be enough to win a place inside U.S. government networks.