US DeepSeek ban for federal agencies now has deadlines and limits

The Department of Homeland Security has issued an order banning DeepSeek products and services from federal civilian agencies. The notice, made public in the Federal Register, says agencies must remove DeepSeek and its affiliated products from their IT systems. It also blocks civilian agencies from buying DeepSeek products and services. The order covers third-party products and services built using DeepSeek, not just DeepSeek's own offerings.

Agencies have 60 days to purge their systems. Those with a compelling justification can ask the Cybersecurity and Infrastructure Security Agency (CISA) for an exemption by written request. DHS cited several concerns: DeepSeek's servers are located in China, Chinese authorities could potentially access private company infrastructure and sensitive U.S. data, and the models may contain security vulnerabilities. The authority comes from the Federal Acquisition Supply Chain Security Act of 2018, following a review by the Federal Acquisition Security Council.

The limits matter as much as the ban. The order applies to civilian executive branch agencies, including Treasury, State, and Justice. It does not cover consumers, state and local governments, or Defense Department national security systems. DeepSeek's open-source models remain widely available in countries including the United States, and they are mostly free to use.

There are open questions. The signed order is dated July 24, but it was only made public in October, and one account says the 60-day exception window closed before the public notice appeared. It is also unclear how the order applies to finetuned models or models trained on DeepSeek's outputs. China's Foreign Ministry said it opposes overstretching the concept of national security and will safeguard the rights of Chinese companies.

Our take

This looks like the first time a frontier AI model has been treated as a supply-chain security threat, and that precedent may outlast this specific order. The hard part is enforcement: policing open-weight models buried inside other software is a murky compliance problem.