DHS Orders DeepSeek Removed From Federal Civilian Agencies Under Supply-Chain Law
A DHS exclusion order, signed June 26 and published in the Federal Register this week, bars civilian federal agencies from procuring DeepSeek products and requires removal of deployed versions, citing multiple threat vectors under the 2018 Federal Acquisition Supply Chain Act.
The Department of Homeland Security has ordered all federal civilian agencies to stop acquiring DeepSeek products and services and to strip out versions already deployed, per reporting from Reuters and CyberScoop published October 19. The order rests on the Federal Acquisition Supply Chain Act of 2018, following a Federal Acquisition Security Council review, and carries the mechanics of that regime: Secretary Mullin’s exclusion order was signed June 26, meaning the ban has been operating for months before this week’s Federal Register publication (docket 2026-21447) made it broadly visible. Mullin’s stated rationale: DeepSeek models present “multiple threat vectors” to civilian agency IT systems, chiefly the risk that government information reaching China, “either independently or in combination with publicly available data,” could yield intelligence on US persons or programs.
What the order actually requires
Three obligations, in procurement language that has teeth. Civilian agencies are barred from procuring IT products or services from parent company Hangzhou DeepSeek Artificial Intelligence and its affiliates. Agencies must remove any DeepSeek versions already in their systems, with interim guidance to add “enhanced logging and controls on data exfiltration for the relevant systems” while removal proceeds. And exceptions exist but are deliberately hard: a waiver request must be submitted 60 days in advance with a “compelling justification,” such as proof that no alternative can fulfill a mission-critical function. The third-party angle matters most in practice: the ban reaches services that use DeepSeek, not just DeepSeek’s own apps, which sweeps in a wide swath of AI tooling that resells Chinese open-weight models under other brands. Defense and intelligence agencies sit outside this order only because their restrictions were already stricter.
The enforcement pattern is consolidating
What makes this order a milestone rather than another headline is the mechanism. Individual agencies had been blocking DeepSeek at their own discretion since early 2025, and the January 2026 debate was about penalties targeting DeepSeek’s access to US technology. This order instead runs the problem through the Federal Acquisition Security Council, the supply-chain body built for hardware like Huawei and Kaspersky, and produces a single government-wide exclusion with a public docket. It is the first time a Chinese AI model has gone through that pipeline, and the template generalizes: the same docket format, the same 60-day exception process, and the same “affiliates” sweep could process Kimi, Qwen, or GLM if the council finds similar threat vectors. Note the irony of the timing: DeepSeek open-sourced its production inference engine with equal NVIDIA and Huawei support the same day the ban surfaced, and Washington’s distillation accusations against the company, documented in Anthropic’s September threat intelligence report, remain the unaddressed backdrop to the security rationale.
What to watch
Three threads determine whether this is a one-model action or the start of a procurement regime for AI. First, scope creep by analogy: procurement lawyers at other agencies and allied governments will read the docket as precedent, and any model with similar data-routing characteristics is now arguable under the same authority. Second, the compliance burden lands on third-party AI vendors, who must now certify their supply chains down to the weights and inference provider, a requirement nobody’s SOC 2 covers; expect a market for AI provenance attestations to appear quickly. Third, the retaliation loop: Beijing has previously responded to US AI restrictions with export-control moves of its own, and a procurement ban under supply-chain law gives China a symmetrical justification for excluding US models from its own government market. The DeepSeek brand survives fine in commercial markets; the question is whether government procurement becomes the wall that splits the global AI market in two.
Get Insanely Good at AI
The book for developers who want to understand how AI actually works. LLMs, prompt engineering, RAG, AI agents, and production systems.
Keep Reading
What Is Mixture-of-Experts (MoE) in AI?
MoE models have a trillion parameters but only activate a fraction per token. How expert routing works, why it matters for cost, and which major models use it.
US Proposes an AI Incident Alert System to China Ahead of the Trump-Xi Summit
Treasury Secretary Scott Bessent said the US proposed a notification mechanism with China for AI safety incidents that reach the national security level, during Sunday's talks with Vice Premier He Lifeng in Manhattan.
DeepSeek Open-Sourced Its Production Inference Engine, and Huawei Gets Equal Billing
DeepSeek released DeepInfer on October 20 under Apache 2.0: the inference framework behind its production services, with NVIDIA SM100 and Huawei Ascend 950 support given equal footing, optimized kernels, and speculative decoding built around confidence scheduling.
Anthropic's Draft AI Safety Bill Would Only Allow Blocking Models Above 50 Deaths or $1B in Damages
Anthropic quietly pitched US lawmakers a draft AI safety bill, reported October 16 by Politico, that would let government block a model's release only for foreseeable, imminent risks of killing more than 50 people or causing over $1 billion in damages, with the Commerce Department as sole gatekeeper.
DeepSeek Refreshed V4.1-Flash Eight Days Later: 763B Parameters, MIT License, and a KV-Cache Squeeze
DeepSeek shipped DeepSeek-V4.1-Flash-1017 on October 17, a checkpoint refresh with large agent-benchmark jumps (Terminal-Bench 4.0 up from 31.2 to 42.3), MIT licensing, 1M context, and a research focus on pushing KV cache compression.