DeepSeek deep-sixed at all US federal civilian agencies

The US Department of Homeland Security on Monday issued an order barring federal civilian agencies from using products from Chinese AI lab DeepSeek and requiring them to remove existing implementations, covering "all DeepSeek and DeepSeek-affiliate products, and third-party products and services developed using DeepSeek implementations." The order's scope is unclear on whether "developed using" covers API access to DeepSeek models or apps built with DeepSeek assistance, and it leaves open how agencies should handle AI gateways such as OpenRouter and Vercel AI Gateway that can route requests to DeepSeek models by default. The directive follows earlier DeepSeek bans at the Defense Information Systems Agency in January 2025, the US Navy, NASA and the US House of Representatives, and the February 6, 2025 introduction of the No DeepSeek on Government Devices Act by Representatives Darin LaHood (R-IL) and Josh Gottheimer (D-NJ), which never advanced out of committee.

DeepSeek deep-sixed at all US federal civilian agencies The Register https://www.theregister.com Order disallowing use of Chinese AI service is unhelpfully vague and broad The US Department of Homeland Security on Monday issued an order directing federal civilian agencies not to use technology products from Chinese AI lab DeepSeek https://www.machinebrief.com/compare/llama-4-vs-deepseek-r1 , and to remove any existing implementations. The order https://public-inspection.federalregister.gov/2026-21447.pdf PDF is broad, as it requires agencies to ditch "all DeepSeek and DeepSeek-affiliate products, and third-party products and services developed using DeepSeek implementations." Complying will be tricky as it is not clear which third-party products and services have been developed using DeepSeek's wares. Nor is it clear whether "developed using" means providing ongoing access to DeepSeek models via API or simply refers to any app vibe-coded with the help of DeepSeek. It also leaves open the question of AI gateways. Services such as OpenRouter and Vercel AI Gateway make DeepSeek's models available through APIs. And those using gateway services may find themselves using DeepSeek models without meaning to do so. OpenRouter, for example https://openrouter.ai/docs/guides/routing/provider-selection , "routes requests to the best available providers for your model." It selects what the service deems to be the best fit provider of AI services. Vercel AI Gateway does the same by default https://vercel.com/docs/ai-gateway/models-and-providers/provider-options . So federal civilian agencies could be using DeepSeek models inadvertently, if they haven't taken steps to constrain their AI providers. A DHS spokesperson did not immediately respond to a request for comment. US officials have contemplated a ban on DeepSeek since the China-based AI lab released DeepSeek-R1 in early 2025. On February 6, 2025, US Representative Darin LaHood R-IL and US Representative Josh Gottheimer D-NJ introduced https://lahood.house.gov/2025/2/lahood-gottheimer-introduce-legislation-to-counter-ccp-s-deepseek-ai-software the No DeepSeek on Government Devices Act. The lawmakers characterised DeepSeek as a national security threat because its AI models purportedly capture data on US citizens and provide it to the Chinese government. The bill never made it out of committee, but various government agencies followed through. The US Defense community and lawmakers have already exiled DeepSeek. The Defense Information Systems Agency DISA started blocking DeepSeek https://www.bloomberg.com/news/articles/2025-01-30/pentagon-workers-used-deepseek-s-chatbot-for-days-before-block on the Pentagon’s IT networks in January 2025. The US Navy https://www.cnbc.com/2025/01/28/us-navy-restricts-use-of-deepseek-ai-imperative-to-avoid-using.html , NASA https://www.cnbc.com/2025/01/31/nasa-becomes-latest-federal-agency-to-block-chinas-deepseek.html , and the US House of Representatives https://www.axios.com/2025/01/30/house-congress-bans-deepseek-ai took similar steps. So did various US states. While the DHS announcement has taken a long time to reach civilian agencies, its timing may be a boon for American AI companies which, like Anthropic https://www.machinebrief.com/glossary/anthropic , hope to soon go public. That's because open weight https://www.machinebrief.com/glossary/weight models represent competition, and those from China have proven particularly popular both in terms of capabilities and cost. Government agencies that cannot deploy such models and the taxpayers who fund those agencies may be stuck paying the prices set by favored US model makers. Concerns about the security and privacy of DeepSeek's iOS mobile app surfaced shortly after DeepSeek-R1 debuted https://www.nowsecure.com/blog/2025/02/06/nowsecure-uncovers-multiple-security-and-privacy-flaws-in-deepseek-ios-mobile-app/ . Several months later, a December 2025 report https://www.sciencedirect.com/science/article/abs/pii/S2666281725001684 on the DeekSeek mobile app for Android and iOS cited similar concerns, including plaintext information leaks, links to servers run by Chinese tech firms, and shortcomings like lack of SSL pinning. These issues, however, are mainly mobile security problems that afflict many apps, particularly those that integrate third-party advertising SDKs from companies based outside the US. They don't really address the additional risks posed by the capture of text prompts and responses, or how local secrets and other sensitive information can be ingested by AI agents. Katie Moussouris, founder and CEO of Luta Security https://www.lutasecurity.com , told The Register in an email that while there are legitimate national security reasons to keep DeepSeek's app and servers away from government systems, the way this particular order is written exceeds that requirement in ways that aren't practical. "Banning DeepSeek's own app and API is workable," said Moussouris. "The problem is the ban on 'third-party products and services developed using DeepSeek implementations.' The order doesn't say what 'developed using' means. It could technically cover any model derived from DeepSeek and more importantly, any software whose developers used DeepSeek to write or test code. An AI software bill of materials doesn't record which tools developers used, so it would be no real help. Every vendor would have to attest that nothing it relies on ever touched DeepSeek. There's no practical way to verify that." Moussouris also took issue with the breadth of the order, noting that applying the rules to US-hosted versions of DeepSeek's models goes too far. "The risks the order cites are mostly about data flowing to DeepSeek's servers in China, its data collection, and its app," she said. "US hosting avoids most of those. "For the model itself, the order says only that it may contain malware or vulnerabilities, which is true of all software. NIST found DeepSeek more vulnerable to prompt injection than US models https://www.nist.gov/news-events/news/2025/09/caisi-evaluation-deepseek-ai-models-finds-shortcomings-and-risks , though every model has that problem. NIST also found it easier to jailbreak https://www.nist.gov/document/caisi-evaluation-deepseek-ai-models-report , but jailbreak https://www.machinebrief.com/glossary/jailbreak guardrails https://www.machinebrief.com/glossary/guardrails often get in the way of defenders." ® Get AI news in your inbox Daily digest of what matters in AI. Key Terms Explained Anthropic https://www.machinebrief.com/glossary/anthropic An AI safety company founded in 2021 by former OpenAI researchers, including Dario and Daniela Amodei. Guardrails https://www.machinebrief.com/glossary/guardrails Safety measures built into AI systems to prevent harmful, inappropriate, or off-topic outputs. Jailbreak https://www.machinebrief.com/glossary/jailbreak A technique for bypassing an AI model's safety restrictions and guardrails. Weight https://www.machinebrief.com/glossary/weight A numerical value in a neural network that determines the strength of the connection between neurons.