DeepSeek banned from U.S. federal civilian agencies under new DHS order

Commenters focused less on the hosted DeepSeek app than on the order’s reach into self-hosted open-weight models and downstream products. Many saw a real distinction: sending agency prompts to servers under Chinese jurisdiction creates an obvious data-access risk, but banning local inference means the government is treating the weights themselves as potentially hostile. That claim drew both technical caution and deep skepticism. Several readers noted that neural-network weights are not auditable in the way source code is; a model could conceal a trigger that makes a tool call, leaks data, or retrieves a malicious payload only under rare conditions. The cited proof-of-concept made that threat feel more concrete than a generic China-data-residency argument. Still, the prevailing mood was suspicious of a DeepSeek-only prohibition. Readers argued that the order does not identify evidence of poisoned weights, while its broad language could deter contractors from using any product built around DeepSeek, even entirely local deployments. Some viewed this as a strategically selective response to an unusually capable, cheap Chinese competitor, with fear likely to spread to other Chinese open models once they matter commercially. Others pushed back that governments routinely exclude technology from geopolitical adversaries, and that open weights do not eliminate the hard problem of discovering latent malicious behavior. There was also appreciation for local models’ ability to be post-trained for agency-specific capability and alignment, even if that does not prove the original release safe.