US bans DeepSeek across federal agencies, putting AI supply chains under scrutiny
For business leaders, the order points to a bigger challenge: finding AI models built into products, tracking where data goes, and managing risks from outside suppliers.
DeepSeek has suffered a major setback in the US, even as it continues to develop new technology and keep up with, or sometimes perform better than, competing AI models from American research labs.
The US Secretary of Homeland Security has issued an order banning all DeepSeek products and services from federal agencies. Government organizations must remove them from their IT systems within 60 days. Agencies that have a strong reason to keep using them can ask for an exception in writing from the Cybersecurity and Infrastructure Security Agency (CISA).
The Department of Homeland Security (DHS) gave several reasons for its concerns about the AI model hosted in China. These included data sovereignty, weak Chinese laws, and security weaknesses.
Erik Avakian, a technical counselor at Infotech, understands these concerns. “When organizations use any type of hosted AI service, they may send intellectual property, confidential business information, or other sensitive data into an environment where they have very little ability to see or control what happens to it,” he said.
Government concerns
According to a document published in the US Federal Register, federal civilian agencies will be banned from buying technology developed by China-based Hangzhou DeepSeek Artificial Intelligence Co. The ban also covers related products and third-party tools or software built using DeepSeek technology.
The civilian agencies affected include the Departments of Treasury, State, and Justice. The ban does not apply to private consumers, the general public, state or local governments, or national security systems used by the Department of Defense (DoD).
“DeepSeek’s models present multiple threat vectors,” the order stated. First, the company's servers are located in the People's Republic of China (PRC). Chinese law gives government cyber and intelligence agencies “extensive authority” to access private computer systems and infrastructure.
The order also claimed that DeepSeek collects personally identifiable information (PII) when people create accounts. It may also track users and create digital fingerprints that can identify them. When users send prompts and receive answers, sensitive data may be transferred to China. This could provide “actionable intelligence” about US programs or individual users.
The order further stated that DeepSeek models may contain malware or other security weaknesses. DHS pointed to security research published in February 2025. The research found several serious security and privacy weaknesses in iOS mobile apps.
Until the ban fully takes effect, DHS advised agencies to reduce risks by watching data traffic coming into and leaving their systems, keeping more detailed records, using stronger controls to prevent data from being sent out without permission, and detecting and flagging unusual activity.
Agencies can request a written exception to the ban, but they must provide a “compelling justification.” Reasons could include the effect of the ban on essential government work or other national interests, such as national security reviews, investigations, or security agreements.
Security problems remain widespread
The ban comes just as DeepSeek is making DeepInfer, its own AI inference framework, open source. DeepInfer powers DeepSeek-V4.1-Flash. The company says it can process large amounts of work quickly with little delay. Its features include high-performance kernels, an optimized inference engine, structured outputs, and speculative decoding.
DeepSeek said its production components for encoding and decoding tokens are designed to follow the processing logic behind the DeepSeek API. The company said these components would provide “the same API behavior to community deployments.”
“Open models. Open kernels. Open inference,” DeepSeek posted on the social media site X. “From private deployments to public clouds, we’ll keep working with the community to make large-scale inference more efficient.”
This is not DeepSeek's first cybersecurity setback. In January 2025, the company exposed more than one million log entries through a database that was not properly secured. The records included chat histories, API secrets, and other sensitive information.
Avakian said DeepSeek reportedly fixed the exposure. However, the incident raised questions about the platform's overall security, especially for organizations thinking about sharing their own sensitive business information with it.
He said this does not mean that every DeepSeek model is “inherently malicious.” But important questions remain about data privacy, whether Chinese laws could allow access to information, the security of the wider AI ecosystem, and whether DeepSeek models can be tricked into ignoring their safeguards or manipulated.
“Taken together, [these issues] raise legitimate concerns about whether organizations can adequately understand and manage the risks involved,” Avakian said.
Businesses should carefully assess the risks
More businesses are exploring DeepSeek, Avakian said, although most publicly reported use has been in China. Some DeepSeek models are also available through platforms such as Microsoft Azure and AWS.
“There’s genuine enterprise interest in the technology, as organizations look for opportunities to improve productivity, offer new capabilities, or reduce costs,” Avakian said. Possible uses include summarizing documents, doing internal research, and helping with coding.
DeepInfer could also help lower costs, but businesses should weigh its possible benefits against security risks, he said. The framework has only just been released. Organizations need to test its claimed efficiency gains in their own systems instead of assuming it will automatically reduce their operating costs.
Avakian advised organizations to start with uses that carry relatively little risk. They should preferably use “public, synthetic, or appropriately sanitized information” in a separate, controlled test environment that they manage themselves.
The first question to ask is: “What is the business problem we’re trying to solve?” Avakian said. Next, organizations should decide whether DeepSeek is the best way to solve that problem, check that security controls work, and add more uses based on how much risk they are willing to accept. Strong governance and risk assessment are essential, he said. Organizations need to understand what they are adding to their systems, test its security, limit what it is allowed to do, and make sure they can monitor its activity.
“Approach this like any other software supply chain decision,” Avakian advised. “The fact that something is open source, inexpensive, or technically impressive doesn’t automatically make it appropriate for enterprise use.”
Effects on the wider AI ecosystem
Avakian said it was important that the order also covered third-party products and services built using DeepSeek technology. Organizations may be using this technology indirectly through suppliers, software platforms, or other services without knowing what is running behind the scenes. “These are classic supply chain issues,” he said.
Avakian said it is no longer enough for businesses to know which AI applications their employees use. They also need to know which AI models, frameworks, supporting software components, and outside services are built into the products they buy and use. He advised businesses to review their lists of AI tools, consider updating their purchasing rules, and improve how they manage risks from third-party suppliers.
“Ultimately, enterprises leveraging these tools need to understand not only the AI technologies they’re intentionally deploying, but also the technologies they’re inheriting through third parties,” he said.