Federal Agencies Told to Remove DeepSeek, Including Indirect Use Through Other Vendors
DHS has ordered federal civilian agencies to exclude DeepSeek products from their information systems and to remove existing installations. Engadget The directive applies to federal civilian executive branch information systems.
The order also covers indirect use. It includes third-party services that use the DeepSeek AI model, in addition to DeepSeek products themselves. Yahoo News The notice also disallows DeepSeek from participating in federal government contracts. Engadget
DHS cited two concerns. The department said DeepSeek use could potentially result in transfer of sensitive data about U.S. government employees and vendors to the Chinese government. It also said DeepSeek's models could contain malware or other security vulnerabilities. Engadget
The order came a couple of days after DeepSeek released AI model V4.1-Flash-1017, which the company described as "smarter, faster, more efficient." Engadget In April 2026, DeepSeek had previewed a new AI model described as closing the gap with frontier models, the term for the most capable systems available. TechCrunch Australia had banned DeepSeek from all government devices over concerns that it posed security risks. Reuters
Looking at what this means for technology teams, the difficult part is the indirect access. A direct install is straightforward to find and remove. It gets harder when the model arrives through another vendor, a software kit that builds the model into an app, a server that serves up AI answers, or an internal pilot that calls an outside service through an API, a standard software connection. Purchase records rarely show that full chain. Logs of outbound internet traffic, gateway inventories, and software bills of materials, which list the parts inside software, do.
The broader context here is that DHS named two distinct risks, and they call for different controls. Data transfer risk is about where prompts, meaning user inputs, completions, meaning AI outputs, tracking data, and account details go, and under what legal jurisdiction and storage rules. Malware or vulnerability risk is about what runs locally, which model files and programs are trusted, and how updates are checked. In practice, that means isolating AI traffic, using allowlists of approved services, checking files against known fingerprints, and requiring proof of origin for any model used in sensitive work.
In my view, orders of this kind will push federal IT and its contractors toward stricter proof of model origin as standard practice. That is tedious work. It is also useful work. Clear exclusion lists and contract language reduce guesswork for security leaders and platform owners who otherwise must judge each fast-moving release case by case. Over the long arc, that plumbing should make adoption easier, because teams can say yes faster to systems with known lineage, audit trails, and support boundaries.
