US Orders Civilian Agencies to Remove DeepSeek, Extending Ban to Outside Suppliers

DeepSeek ban reaches beyond the chatbot

The US Department of Homeland Security has ordered federal civilian agencies to remove DeepSeek technology from their information systems and stop acquiring products and services built with it, extending Washington's restrictions on the Chinese artificial intelligence company into the software supplied by outside contractors.

A DHS notice filed with the Federal Register on October 19, 2026, described DeepSeek models as presenting "multiple threat vectors." The department warned that use by government employees or vendors could send sensitive information to China, produce actionable intelligence about Americans or federal agencies, and expose systems to malware or other security weaknesses.

The restriction reaches beyond downloading DeepSeek's app or visiting its website. It covers DeepSeek and affiliated products, along with products and services from other companies developed using DeepSeek implementations. Civilian agencies such as the Environmental Protection Agency, the Treasury Department and the Internal Revenue Service fall within its scope.

The order was reportedly signed on July 24, before its October public disclosure. A reported window of 60 days for requesting exceptions would therefore have closed around September 22. That timing gives government suppliers an immediate compliance problem, while the breadth of the language leaves uncertainty about exactly which software they must remove.

What DHS says agencies must remove

DHS directs civilian agencies to exclude existing DeepSeek technology and stop obtaining covered products and services. The acquisition restriction means an agency cannot simply replace a direct DeepSeek subscription with another company's service that supplies the same underlying model.

An AI model is the software system that generates answers, writes code or performs other tasks. A company can place its own interface around another developer's model, so the name on a government software contract may differ from the name of the technology doing the work.

For procurement teams, the practical task is therefore broader than searching a list of installed apps. Agencies need to understand which models their suppliers use, including models reached through cloud services and subcontractors. A DHS memo issued in July also urged agencies to monitor traffic to and from government websites for potential DeepSeek risks.

The civilian order operates alongside separate rules governing defense and intelligence systems. Specified national security systems fall under other authorities, while Pentagon networks and military users have already faced restrictions on DeepSeek.

Why the government sees a security risk

DHS identifies several ways government use of DeepSeek could create exposure. Sensitive material entered by an employee or vendor could move to China. Information collected across many interactions could also help someone build a useful picture of an agency, its staff or its operations.

The department's concern also extends to software security. Its notice says DeepSeek models could "contain malware or other security vulnerabilities." For an agency using AI to process documents or assist with computer tasks, a weakness can matter beyond the answer displayed in a chat window.

Security examinations of DeepSeek's mobile apps have raised more specific concerns. A December 2025 assessment described information transmitted in plaintext, connections to servers operated by Chinese technology companies and a lack of SSL pinning. Plaintext information lacks encryption during transmission. SSL pinning is an additional check that helps an app confirm it is communicating with the intended server.

Model behavior creates another risk. Tests attributed to the National Institute of Standards and Technology found agents built on DeepSeek R1-0528 followed planted malicious instructions about 12 times as often as agents using GPT-5 and Claude Opus 4. This type of attack, called prompt injection, hides instructions inside material an AI system reads, attempting to redirect it away from its assigned task.

The difficult question of outside software

The phrase covering products "developed using DeepSeek implementations" creates the largest practical uncertainty. It could reach a service that actively sends requests to DeepSeek, a model adapted from DeepSeek's technology, or software whose developers used DeepSeek to write or test code. DHS has not publicly clarified those boundaries in the accounts of the order.

Katie Moussouris, founder and chief executive of cybersecurity company Luta Security, said the direct restrictions are easier to carry out than the provision covering outside products:

Banning DeepSeek's own app and API is workable. The problem is the ban on 'third-party products and services developed using DeepSeek implementations.'

An API, or application programming interface, lets one software system request services from another. Blocking an API connection is a concrete technical step. Establishing which coding assistant a supplier used months earlier requires a different kind of record, which may never have been kept.

Moussouris said an AI software bill of materials, essentially an inventory of a product's AI components, would not record every tool used during development. She described the verification problem in direct terms:

Every vendor would have to attest that nothing it relies on ever touched DeepSeek. There's no practical way to verify that.

Cloud routing can make DeepSeek harder to spot

Services such as OpenRouter and Vercel AI Gateway allow customers to reach multiple AI providers through a common connection. Depending on the model and routing settings, a request can be directed to a provider without the employee making a fresh choice each time.

A government supplier therefore needs to examine configuration settings as well as contracts. Selecting a permitted model explicitly, checking fallback options and reviewing service logs can help identify whether a product reaches DeepSeek through an intermediary.

DeepSeek also releases open weight models. The weights are the numerical files containing what a model learned during training. Other companies can download those files and run the model on their own computers or through a US cloud provider, rather than sending requests to DeepSeek's servers.

Moussouris argued that hosting location matters when assessing the government's stated concerns:

The risks the order cites are mostly about data flowing to DeepSeek's servers in China, its data collection, and its app. US hosting avoids most of those.

The order's broad wording still presents a compliance issue for agencies considering those arrangements. Running a model in the United States changes where information is processed, while the restriction also addresses the technology's origin and its use inside other products.

How the restrictions developed

Federal action against DeepSeek began well before the latest civilian notice. Defense organizations, individual agencies and state governments imposed their own restrictions as the company's models gained users.

• January 2025: The Defense Information Systems Agency began blocking DeepSeek on Pentagon networks. DeepSeek's inexpensive AI assistant also drew worldwide attention that month.
• February 6, 2025: Representatives Darin LaHood and Josh Gottheimer introduced the No DeepSeek on Government Devices Act. The bill did not leave committee.
• During 2025: The Commerce Department restricted DeepSeek on government devices, and the Navy ordered its members not to use it. NASA, the US House of Representatives and states including New York, Tennessee and Virginia also took action.
• July 2026: DHS urged agencies to monitor website traffic for DeepSeek risks. The broader removal order was reportedly signed on July 24.
• October 17, 2026: DeepSeek released V4.1-Flash-1017, its latest model.
• October 19, 2026: DHS publicly disclosed the civilian agency restriction through its Federal Register notice.

Federal scrutiny has also extended to allegations that Chinese AI developers used distillation attacks against leading US models. Distillation involves training one model using another model's answers. US concerns focus on alleged unauthorized extraction of capabilities from proprietary systems, alongside the separate security concerns behind the removal order.

Lower prices explain DeepSeek's appeal

DeepSeek's attraction has been its combination of capable models, downloadable weights and relatively low prices. Those features made it a competitive option for organizations seeking to use AI without relying entirely on the largest American providers.

The October 17 release sharpened that competition. AI evaluation company Artificial Analysis scored V4.1-Flash-1017 at 45, compared with 39 for its predecessor, while its average evaluation cost fell from $0.49 to $0.26 per task. That is a reduction of about 47%, calculated from the two figures.

DeepSeek's own performance tables put the new model's DeepSWE v1.1 software engineering score at 77.0, up from 74.2. Its Terminal-Bench 4.0 score rose from 31.2 to 42.3. These tests examine particular tasks, giving prospective users measures to consider alongside security, reliability and operating costs.

The company's quoted API prices were $0.30 per million input tokens and $1.20 per million output tokens during peak hours, with rates cut in half outside those periods. Tokens are small pieces of text that AI companies use to measure how much information a model receives and generates.

DeepSeek's January 2025 breakthrough had already demonstrated how strongly cheaper AI could affect markets. Losses across semiconductor, power and infrastructure companies exceeded $1 trillion in market value. Nvidia fell about 17%, losing nearly $593 billion in what was then a record loss in a single day for a company.

Government suppliers face overlapping rules

Businesses serving civilian agencies now have to examine their products against the DHS restriction. Defense contractors face a separate layer: Section 1532 of the fiscal 2026 defense authorization law covers DeepSeek and its parent company, High-Flyer.

Federal acquisition clause FAR 52.204-30 also creates compliance duties for contractors subject to it. Those duties reportedly include checking for applicable supply chain orders at least every three months, reporting identified covered technology within three business days and providing remediation information within another 10 business days. A supplier's obligations depend on its contract and the applicable order.

For companies selling both commercial and government services, the immediate business issue is separation. The civilian agency restriction concerns federal systems and purchasing. Suppliers using DeepSeek elsewhere need to identify whether shared infrastructure, subcontractors or development processes bring that technology into government work.

Further restrictions are under consideration. Proposed provisions for the fiscal 2027 defense bill would extend limits to ten additional Chinese AI companies and derivative models. Those proposals remain distinct from the restrictions already in force, but they give suppliers another reason to track the underlying models in their products.

FCC plans another move on Chinese technology

The Federal Communications Commission has scheduled an October 29 vote on a separate proposal concerning laboratories that test electronic devices for the US market. It would bar Chinese and other foreign laboratories in countries that deny reciprocal treatment to US testing labs.

The FCC had already barred laboratories owned or controlled by the Chinese government from testing US electronics in 2025. The commission said many laboratories based in China continued to carry out that work, prompting the broader proposal.

FCC Chair Brendan Carr said 82% of electronics were tested in China, compared with fewer than 4% in US laboratories. Using those figures, China's share was more than 20 times the US share. Testing laboratories assess whether devices meet technical requirements before they enter the market, making their approval role consequential for manufacturers.

The October 29 vote is the next scheduled decision. The DeepSeek order and the laboratory proposal address different technologies, but both extend US scrutiny beyond a finished product to the companies and services behind it. For agencies and vendors, knowing who supplies the underlying technology is becoming a central part of doing government business.

Key Points

• DHS has directed federal civilian agencies to remove DeepSeek products and covered services supplied by other companies.
• The October 19, 2026, public notice disclosed a restriction reportedly signed on July 24.
• DHS cited sensitive data transfers, intelligence risks and software security weaknesses.
• The meaning of products developed using DeepSeek remains a practical concern for suppliers.
• DeepSeek's latest model arrived on October 17 with improved evaluation scores and lower measured task costs.
• Defense contractors face separate restrictions, while broader limits on Chinese AI are being considered.
• The FCC plans an October 29 vote on restrictions affecting foreign electronics testing laboratories.