Every prompt you send to a hosted AI model is a data transfer, and you rarely get to decide where it ends up. That is the uncomfortable core of the US ban on DeepSeek.
The Department of Homeland Security announced a directive on Monday barring Hangzhou DeepSeek from federal civilian agency networks. General Counsel James Percival wrote that the company’s models present “multiple threat vectors.” His main argument is jurisdiction. DeepSeek’s servers sit in China, where law gives government cyber and intelligence bodies broad authority to reach into private company infrastructure. Transferring data there, the directive says, “could produce actionable intelligence on US persons and/or government programs.”
The technical claims are thinner, and DHS offers little detail. It says independent assessments found critical vulnerabilities in the company’s mobile software, plus user fingerprinting risks. Fingerprinting matters because it tracks a device even when you avoid logins and cookies. But we haven’t seen the assessments, so readers should treat those findings as claims until someone publishes them.
The order itself is blunt. Civilian agencies must immediately remove DeepSeek software, affiliated tools and any third-party apps built on its framework. The company comes off approved product lists. Agencies can request exemptions for mission-critical work within 60 days through the Cybersecurity and Infrastructure Security Agency. Until removal is finished, they must tighten network logging, restrict outbound data transfers and watch for anomalous traffic.
That last part is the most useful. The ban on third-party apps built on DeepSeek is an admission that AI supply chains are messy. A tool may look American and still route your queries to a model hosted somewhere else. Procurement lists rarely capture that.
Beijing answered Tuesday. Foreign Ministry spokesperson Mao Ning said China opposes “overstretching the concept of national security and politicizing science, technology and economic and trade issues.” She promised to defend the “legitimate rights and interests of Chinese companies.”
Both sides are performing, and both have a point. China’s intelligence laws are real and worth worrying about. But US companies also hand data to American agencies under their own legal powers, and users have little visibility into either. The ban fits a pattern that began with TikTok restrictions and Huawei equipment bans, where data location and legal reach decide who gets trusted.
For ordinary users, the lesson is simple. Check where a service stores prompts, what it logs, and who can compel access. Don’t paste confidential work, health details or credentials into any hosted chatbot. And if privacy matters, consider running open models locally, where your queries never leave your machine.