US DHS Bans DeepSeek Across All Federal Agencies, Including Third-Party Products

DHS bans DeepSeek across all US federal civilian agencies, including third-party products built on DeepSeek technology. Open-weight becomes a liability.

DHS issues order: DeepSeek banned across all federal agencies

On October 20, 2026, the US Department of Homeland Security (DHS) published an official order banning DeepSeek across all federal civilian agencies. The order doesn't just ban DeepSeek's own products, but also "third-party products and services developed using DeepSeek implementations."

This order was signed by DHS Secretary Markwayne Mullin on July 24, 2026, but was only published in the Federal Register on October 20. This means the ban has actually been in effect since July, only now announced to the public.

Affected agencies: Treasury, State, Justice, and all other federal civilian agencies. The Department of Defense (DoD) already blocked DeepSeek back in January 2025 (DISA, Navy, NASA, House).

Reasons: data, China servers, and malware

DHS cites several reasons:

1. Data stored in China. DeepSeek models store data on servers in China, where national security laws give the Chinese government broad access to domestic company data. Sensitive US government data could "leak" to China.
2. User data collection. DeepSeek models collect personally identifiable information (PII) and "digital fingerprints" of users. This data could be combined with public data to produce "actionable intelligence" about US government personnel.
3. Malware and vulnerabilities. DHS cites a NowSecure report (February 2025) that found security vulnerabilities in the DeepSeek iOS app: unencrypted data transmissions, weak encryption algorithms, insecure local storage, and data sharing with ByteDance (TikTok's parent company).
4. Prompt injection. DHS also notes that all AI models, including DeepSeek, are vulnerable to prompt injection that could result in sensitive data transfer.

Open-weight becomes a liability

This is a big irony. DeepSeek became popular because it's open-weight: free, cheap, and embeddable in other products. But precisely this "open" nature is what makes it "bannable."

The DHS order doesn't just ban DeepSeek itself, but also "third-party products and services developed using DeepSeek implementations." Because DeepSeek is open-weight, models embedded in other products are also banned.

The problem: it's almost impossible to verify. Katie Moussouris, CEO of Luta Security, points out that "there is no practical way to verify" whether a product uses DeepSeek. AI software bill of materials doesn't record which tools developers used. Every vendor would have to attest that nothing they rely on ever touched DeepSeek, and that can't be proven.

Even more complicated: OpenRouter and Vercel AI Gateway do automatic routing. They might choose DeepSeek as the best provider without the user knowing. So federal agencies could be using DeepSeek without knowing it.

Impact: US government AI market closed to Chinese models

Direct implications:

• Vendors using DeepSeek in their products lose the US government market
• Federal agencies must pay more for US models (OpenAI, Anthropic, Google)
• Price competition from Chinese models (which are much cheaper) disappears from the government market

The Register notes: the timing of this ban "may be coincidental" with Anthropic's IPO plans. Open-weight models from China are competition, and governments that can't use them are stuck paying prices set by favored US model makers.

China's response

Chinese Foreign Ministry spokesperson Mao Ning stated Beijing opposes "overstretching the concept of national security and politicizing technology and trade issues." She added: "We will resolutely safeguard the legitimate rights and interests of Chinese enterprises."

China hasn't officially banned US models, but there are indications that Chinese companies are starting to develop models that don't depend on US infrastructure.

Context: not the first time

The DeepSeek ban isn't new. Before the DHS order:

• DISA (Defense Information Systems Agency) blocked DeepSeek January 2025
• US Navy, NASA, US House of Representatives also blocked
• Several US states also blocked
• December 2025 law already targeted DeepSeek for national security systems

This DHS order extends to civilian agencies, including new purchases and already-installed systems. Agencies can request exceptions within 60 days with "compelling justification."

Analysis: open-weight becomes a liability

DeepSeek became popular because it's open-weight: free, cheap, and embeddable. But precisely this "open" nature is what makes it "bannable." This is a big irony.

The fundamental problem: third-party ban is nearly impossible to enforce. There's no way to know whether a product uses DeepSeek as a basis. AI software bill of materials doesn't record which tools developers used.

Even more complicated: OpenRouter and Vercel AI Gateway do automatic routing. They might choose DeepSeek as the best provider without the user knowing. So federal agencies could be using DeepSeek without knowing it.

Dampak bagi vendor dan developer

Vendors selling to the US government must audit their entire stack to ensure no DeepSeek, attest that nothing touches DeepSeek, and switch to US alternatives (more expensive) if they use DeepSeek.

Developers using DeepSeek for prototyping must be careful: if their product is later sold to the government, DeepSeek must be replaced.

Market and industry reaction

The DeepSeek ban has been in effect since July 2026 but was only announced in October. This means federal agencies have already been adapting for three months without the public knowing.

Impact on the AI market:

• US vendors (OpenAI, Anthropic, Google) get a larger government market
• Chinese models lose access to the US government market
• AI prices for federal agencies will likely rise (less competition)
• Startups using DeepSeek must pivot to US models

Some AI startups that use DeepSeek as a backend for their products now have to find alternatives. This includes products already installed in federal agencies.

Comparison with TikTok ban

The DeepSeek ban resembles the TikTok ban: both from China, both involve data, both hard to enforce. But there are differences:

• TikTok: consumer, millions of users, easy to identify
• DeepSeek: enterprise, thousands of products, nearly impossible to identify

The TikTok ban failed because it was hard to enforce and popular among the public. The DeepSeek ban might be easier to enforce for official products, but nearly impossible for third-party products.

Kesimpulan: open-weight jadi bumerang

DeepSeek became popular because it's open-weight: free, cheap, and embeddable. But precisely this "open" nature is what makes it "bannable." Third-party ban is nearly impossible to enforce, and federal agencies are stuck paying prices set by US vendors.

Unanswered question: will this ban encourage US companies to create cheaper open-weight models? Or will it actually strengthen OpenAI and Anthropic's dominance in the government market?

What's clear: open-weight is no longer a pure advantage. It's also a liability that can be used to restrict model usage. This is an important shift in AI geopolitics.

Apa yang perlu diperhatikan

• Ban ini hanya untuk agensi federal civilian, bukan konsumen atau swasta
• Third-party products yang pakai DeepSeek juga kena, tapi verifikasinya hampir mustahil
• OpenRouter/Vercel AI Gateway bisa routing ke DeepSeek tanpa sadar
• Agensi yang butuh exception harus tunggu 60 hari + review CISA
• Moonshot AI (Kimi) dan Alibaba (Qwen) belum kena ban, tapi "worried national security officials"
• DeepSeek ban bisa jadi preseden untuk model China lainnya
• Agensi federal terjebak membayar harga vendor US yang lebih mahal
• Ban berlaku sejak Juli 2026, baru diumumkan Oktobe

Analysis: open-weight becomes a liability

DeepSeek became popular because it is open-weight: free, cheap, and embeddable. But precisely this open nature is what makes it bannable. This is a big irony.

The fundamental problem: third-party ban is nearly impossible to enforce. There is no way to know whether a product uses DeepSeek as a basis. AI software bill of materials doesn't record which tools developers used.

Even more complicated: OpenRouter and Vercel AI Gateway do automatic routing. They might choose DeepSeek as the best provider without the user knowing. So federal agencies could be using DeepSeek without knowing it.

r 2026

Analysis: open-weight becomes a liability

DeepSeek became popular because it is open-weight: free, cheap, and embeddable. But precisely this open nature is what makes it bannable. This is a big irony.

The fundamental problem: third-party ban is nearly impossible to enforce. There is no way to know whether a product uses DeepSeek as a basis. AI software bill of materials doesn't record which tools developers used.

Even more complicated: OpenRouter and Vercel AI Gateway do automatic routing. They might choose DeepSeek as the best provider without the user knowing. So federal agencies could be using DeepSeek without knowing it.

The unanswered question: will this ban encourage US companies to create cheaper open-weight models? Or will it actually strengthen OpenAI and Anthropic's dominance in the government market?

What is clear: open-weight is no longer a pure advantage. It is also a liability that can be used to restrict model usage. This is an important shift in AI geopolitics.

Ringkasan: apa yang perlu diingat

Ban DeepSeek adalah momen penting dalam persaingan AI global. Open-weight bukan lagi keuntungan murni, tapi juga liability. Agensi federal terjebak membayar harga vendor US. Dan ban third-party hampir mustahil ditegakkan. Ini bukan akhir persaingan AI, tapi awal dari era di mana model China dianggap ancaman keamanan, bukan sekadar kompetitor.