DHS orders DeepSeek removed from US civilian agency IT systems and bars future purchases
- The Department of Homeland Security ordered removal of DeepSeek products and third-party services built on DeepSeek from US civilian agency IT systems, and banned agencies from acquiring them, in a notice published Monday.
- DHS says DeepSeek models pose multiple threat vectors, including possible transfer of sensitive data to China and the risk that the models contain malware or security vulnerabilities.
- A July memo made public Monday told federal agencies to monitor traffic flowing in and out of government websites for risks tied to DeepSeek.
- The Commerce Department and the Navy had already banned DeepSeek on government devices, and New York, Tennessee, and Virginia have taken similar steps.
- The order fits a wider Trump administration crackdown on Chinese tech, including an FCC vote on October 29 to bar Chinese labs from testing electronic devices for use in the US.
Hacker News opinions
Thing is, the wording seems to ban use even when the weights are hosted through other providers. "third-party products and services developed using DeepSeek implementations" So you could run it fully locally inside the agency and still be banned. Pretty blunt.
They're also banning local inference with the malware line. Doesn't look like they found the models actively poisoned though.
Swap PRC for U.S., DeepSeek for OpenAI, and it reads the same. Both countries may as well just say they're banning it because the other one is the closest competitor.
DeepSeek is out of this world good and cheap, that's why.
That's not why. We're in an adversarial era and frontier models have unexpected attack vectors. Inspecting the weights doesn't mean you can derive the behavior.
The models may well have malware, but at least the weights are open to inspection. Can't say the same for OpenAI or Anthropic inference.
How do you inspect weights for an intentionally hidden trigger that writes malware? I don't think inspection is practical.
If the weights are open and it's self-hosted, what's the real risk? Isn't banning that mostly regulatory capture dressed up as security?
Weights being open isn't like code being open. You can't easily infer what a model does from its weights, and post-training could plant backdoors that trigger in certain scenarios.
Missed last year's story? CrowdStrike found DeepSeek generates insecure code when you mention Tibet, and subtler flaws when you mention the US.
It's not regulatory capture to remove software from adversaries. Nobody cried capture when Kaspersky got banned.