China's military used DeepSeek for cyber drills, US Congress report finds

Share:
Audio Loading voice…
China's military used DeepSeek for cyber drills, US Congress report finds

Synopsis

A US congressional report has found that China's military specifically procured DeepSeek for cyber attack training drills and sought to run both DeepSeek and Alibaba's Qwen on command systems — while PLA cyber unit personnel made multiple visits to leading AI developer Zhipu. The findings are the most detailed public evidence yet of the Chinese military integrating homegrown commercial AI into defence operations.

Key Takeaways

A US congressional report released 21 November 2026 found China's military specified DeepSeek for a cyber training range supporting 'intelligent attack' and 'intelligent penetration'.
A January 2026 procurement notice sought servers running both DeepSeek and Alibaba's Qwen for an 'intelligent command information system'.
People associated with the PLA Cyberspace Force's 1st Bureau visited Zhipu AI's Beijing HQ multiple times; the most recent visit was in August .
Four Chinese AI model families — DeepSeek , Qwen , Kimi , and MiniMax — were specified for an isolated internal military network in a May procurement request.
US firms Airbnb and DoorDash told investigators they used Chinese AI models but had no commercial agreements with Chinese developers.
Congressional chairmen called for broader federal bans on Chinese AI in government contracts and critical infrastructure.

China's military has specified DeepSeek for a cyber training facility and sought servers capable of running both DeepSeek and Alibaba's Qwen for a command system, according to procurement records cited in a US congressional report released on 21 November 2026. The findings, drawn from military purchasing documents, raise fresh concerns in Washington about the intersection of Chinese commercial AI and state military capability.

Key Findings from the Congressional Report

The joint investigation was conducted by the House Select Committee on China and the House Homeland Security Committee, led by Republican chairmen John Moolenaar and Andrew Garbarino. The committees examined military procurement records alongside documents from American companies that use some of the same Chinese AI model families.

A procurement notice dated 10 July specified DeepSeek for a military cyber range — a controlled training environment where personnel practise on live computer systems. The notice called for the model to support 'intelligent attack' and 'intelligent penetration', according to the report. Critically, the committees did not identify any live military operation in which DeepSeek had been deployed.

Separately, a January 2026 notice sought servers for an 'intelligent command information system' that would support development work using both Qwen and DeepSeek. Military interest in commercial AI models was traced back to at least October 2024, when a procurement notice called for Zhipu's GLM models to run on a military office network.

Visits to AI Developer Zhipu

The two committees also reported that individuals associated with the 1st Bureau of the People's Liberation Army (PLA) Cyberspace Force had visited the Beijing headquarters of Zhipu, one of China's leading AI developers, on several occasions. The most recent visit was recorded in August. The committees stated they had 'moderate confidence' that these visits indicated a communication channel between the unit and the company. The report gave no account of what the visitors discussed.

The bureau's responsibilities are understood to include signals intelligence, cyber espionage, and electronic warfare, according to the report.

Scope of Military AI Procurement

A May request for proposals set out a broader condition: four model families — DeepSeek, Alibaba's Qwen, Moonshot AI's Kimi, and models from MiniMax — had to function entirely on an internal military network without internet access. The project envisaged using them to develop software and AI agents capable of carrying out autonomous tasks.

The military also sought AI for non-combat functions. A 2 July notice called for DeepSeek or Qwen to run locally to help generate medical records, assess their quality, and assist with clinical decisions.

Notably, the use of open-weight models — which allow buyers to download the model's trained parameters and run them on isolated hardware — means a military purchaser can deploy a commercial AI system entirely offline, with no data flowing back to the original developer.

US Companies Also in the Frame

The congressional inquiry, which began in April with letters to Airbnb and Anysphere (the company behind the Cursor coding platform), later expanded to include DoorDash in July. Airbnb told investigators it ran customised Qwen models on US-based infrastructure for customer support. Both Airbnb and DoorDash said they had neither paid a Chinese AI developer nor entered into a commercial agreement with one.

What the Committees Are Demanding

The committee chairmen have urged Washington to extend federal restrictions beyond DeepSeek alone and to bar Chinese-developed AI models from federal contract work and high-risk functions in critical infrastructure. They also called for a transparency standard that would require suppliers to disclose the underlying AI model when it is incorporated into a product sold under a different brand name.

The report marks one of the most detailed public accounts yet of the Chinese military's use of domestically developed commercial AI — and signals that legislative pressure to tighten AI supply-chain oversight in the US is likely to intensify in the coming months.

Point of View

The developer loses visibility over how it is used, making export controls and usage audits nearly impossible to enforce. Washington's focus on DeepSeek as a single target may be too narrow; the report's own evidence shows the PLA treats multiple Chinese model families as interchangeable tools. The harder question for US policymakers is whether restricting Chinese AI in federal contracts is sufficient when the same models are running on private US infrastructure — including, as the inquiry found, at consumer-facing platforms like Airbnb.
NationPress
21 Nov 2026

Frequently Asked Questions

What did the US congressional report find about China's military and DeepSeek?
The report found that China's military procurement records specified DeepSeek for a cyber training range designed to practise 'intelligent attack' and 'intelligent penetration' techniques. The committees did not identify any live military operation using DeepSeek, but traced military interest in commercial AI models back to at least October 2024.
Which US committees published the report and when?
The joint report was published on 21 November 2026 by the House Select Committee on China and the House Homeland Security Committee, led by Republican chairmen John Moolenaar and Andrew Garbarino respectively.
What is the significance of the PLA visits to Zhipu AI?
People associated with the PLA Cyberspace Force's 1st Bureau visited Zhipu's Beijing headquarters on multiple occasions, with the most recent visit in August. The committees said they had 'moderate confidence' this indicated a communication channel between the cyber unit and the company, though the content of those discussions was not reported.
Why can open-weight AI models like DeepSeek be used by militaries without detection?
Open-weight models allow a buyer to download the trained model parameters and run them on isolated hardware with no internet connection, meaning data never flows back to the original developer. This makes it technically straightforward for a military to deploy a commercial AI model on a closed internal network, beyond the visibility of the model's creator or foreign governments.
What policy actions are US lawmakers seeking following this report?
The committee chairmen urged Washington to broaden federal restrictions beyond DeepSeek and exclude Chinese-developed AI models from federal contracts and high-risk critical infrastructure roles. They also called for a transparency standard requiring suppliers to disclose the underlying AI model when it is repackaged and sold under a different brand name.
Nation Press
The Trail

Connected Dots

Tracing the thread behind this story — newest first.

8 Dots
  1. Latest 1 month ago
  2. 1 month ago
  3. 2 months ago
  4. 3 months ago
  5. 4 months ago
  6. 5 months ago
  7. 5 months ago
  8. 5 months ago
Google Prefer NP
On Google