The US Department of Homeland Security has ordered all federal civilian agencies to stop using and remove DeepSeek AI products — effective immediately. The directive, published Monday in the Federal Register, bans DeepSeek’s own apps and APIs and “third-party products and services developed using DeepSeek implementations.” It applies to agencies including State, Commerce, Health and Human Services, and the Environmental Protection Agency.
Defense agencies were already barred from DeepSeek use: DISA began blocking it on Pentagon networks in January 2025. NASA, the Navy, and the House of Representatives followed suit. Several US states have done the same. The order does not name a compliance deadline, nor does it clarify how agencies should verify third-party compliance.
A DHS spokesperson declined to comment when asked by The Register. The ban comes nearly two years after DeepSeek-R1 launched in early 2025, a model that stunned the industry with high benchmark scores and low inference costs. Its release triggered a $1 trillion US tech stock selloff and prompted bipartisan lawmakers to introduce the No DeepSeek on Government Devices Act in February 2025.
That bill stalled in committee. But the administration moved anyway. Homeland Security Secretary Markwayne Mullin, in a July memo made public only on Monday, cited DeepSeek’s account creation process as a risk: it collects personally identifiable information (PII) and may digitally fingerprint users. He warned that data could flow to Chinese government entities.
The order also alleges DeepSeek models “may contain malware or vulnerabilities”, referencing NIST findings that DeepSeek is more susceptible to prompt injection and jailbreaking than leading US models. Those flaws are real, but NIST noted they’re endemic across the industry. Security expert Katie Moussouris of Luta Security called the third-party clause unworkable.
“An AI software bill of materials doesn’t record which tools developers used,” she told The Register. “Every vendor would have to attest that nothing it relies on ever touched DeepSeek. There’s no practical way to verify that.” She added that applying the ban to US-hosted versions of DeepSeek models, where data never leaves American soil, goes beyond the stated security rationale.
Complicating enforcement: AI gateways like OpenRouter and Vercel AI Gateway route requests to DeepSeek by default unless explicitly restricted. Agencies relying on those services could be using DeepSeek unknowingly. Mobile app concerns also factored in. A December 2025 report flagged plaintext leaks, connections to Chinese servers, and missing SSL pinning in DeepSeek’s Android and iOS apps, issues common in ad-heavy consumer apps, but unacceptable on government devices.
The broader context is tightening. The US government has accused DeepSeek of “illicit distillation”, using outputs from US models like Claude or Gemini to train its own. The NSA, CISA, and FBI jointly named DeepSeek in a recent letter accusing Chinese firms of industrial-scale distillation campaigns. Treasury Secretary Scott Bessent threatened sanctions, calling it “IP theft.” Meanwhile, Beijing itself is moving to rein in overseas AI deployments: Chinese regulators proposed export controls on 23 domestic AI technologies earlier this month, citing national security.
For federal IT teams, the immediate task is triage: identify DeepSeek dependencies, audit gateway configurations, and assess vendor attestations, all without clear definitions or timelines. For US AI startups, the ban removes a low-cost competitor from government procurement pipelines, a quiet boost for Anthropic, Cohere, and others eyeing federal contracts ahead of IPOs.
The order takes effect now. Enforcement details, and whether it spreads to contractors or state-level procurement, remain unannounced.
ADFiled by The AI Desk
Models, assistants and the companies and chips behind them, reported from what was released and what was claimed, with the difference kept clear.
More from this desk →
Be the first to comment
Join the argument. No password, just your email or a passkey.