A letter from Secretary of Homeland Security Markwayne Mullin dated June 26 ordered civilian federal agencies to exclude DeepSeek AI products from their IT systems, indicating the ban has been in place for months.
The notice will appear Tuesday in the Federal Register. It references the Federal Acquisition Supply Chain Act and a review conducted by the Federal Acquisition Security Council. The directive tells agencies to keep DeepSeek out of future purchases. It also calls for the removal of any DeepSeek versions already installed.
Agencies can ask for an exemption from the ban. They must file the request at least sixty days ahead of time. They need to show a strong reason, for example that they cannot carry out a critical mission. A letter from Secretary of Homeland Security Markwayne Mullin dated June 26 ordered the exclusion of DeepSeek products.
The note shows the restriction has existed for a number of months. Mullin said the models introduce a number of threat paths for civilian government computer systems. “multiple threat vectors” He noted that the models keep their data on servers located in China.
Chinese national security rules allow the state to seek and get that information from Chinese firms. He said the models, when hosted by the company, gather personal data from users. They also follow users’ online activity to create digital fingerprints. He observed that, like other large AI systems, DeepSeek works through a prompt‑and‑response method.
This approach could lead to sensitive information being sent to Chinese authorities. He cautioned that moving the data to China might yield useful intelligence. Such intelligence could involve details about individuals or government initiatives in the United States. He urged agencies that drop DeepSeek to strengthen their security measures.
This should involve watching network traffic for unusual patterns. It also means keeping better logs of data leaving the network and spotting odd behavior. DeepSeek stands as the sole Chinese artificial intelligence model specifically mentioned in the DHS directive. Officials also worry about other systems such as Moonshot AI and Alibaba’s Qwen.
The directive further bars any third‑party service that relies on DeepSeek technology from use on government computers. A memorandum from the Department of Homeland Security issued in July appeared on Monday. It earlier told agencies to watch government websites for possible dangers from DeepSeek use. Before the DHS directive, the Department of Commerce had already placed limits on DeepSeek use.
The Navy had also imposed restrictions on DeepSeek earlier. The states of New York, Tennessee, and Virginia had similarly limited DeepSeek access before the order. The Federal Communications Commission will hold a vote on October 29. The vote will consider prohibiting Chinese labs from testing electronic gear meant for U.S. markets.
DeepSeek’s main office lies in Hangzhou. Its open‑source model first gained broad notice on January 27, 2025. Development costs stayed under five million dollars. On the day of its release, NVIDIA’s share price dropped sixteen percent.
This decline wiped out roughly six hundred billion dollars of market value. In April 2025, the House China‑focused committee claimed DeepSeek might be transmitting American user information to China. The committee said the model relied on about sixty thousand NVIDIA processors. Twenty thousand of those chips were subject to export restrictions.
The V4‑Pro preview version of DeepSeek came out in April 2026. Its price was roughly ninety‑eight percent less than that of OpenAI’s GPT‑5.5 Pro. The model is offered at no cost under the MIT license.
Recently DeepSeek put out a new model optimized for efficiency. This model is called the V4.1‑Flash.
Join the discussion