DHS bans DeepSeek across US federal civilian agencies

The US Department of Homeland Security (DHS) has made public an order that bars federal civilian agencies from buying and using products from DeepSeek, the Chinese artificial intelligence lab (Hangzhou DeepSeek Artificial Intelligence Co., Ltd). The order, reported by CyberScoop, The Register, BleepingComputer, DataBreaches.net and heise, appeared on Monday 19 October 2026 in the Federal Register’s public inspection area and was officially published the following day.

According to CyberScoop, the document includes a letter from Secretary of Homeland Security Markwayne Mullin dated 26 June, which means the exclusion was decided months ago. The measure relies on federal supply chain security rules for procurement and on a review by the Federal Acquisition Security Council.

What the order covers

The scope is broad. It covers:

• all products and services from DeepSeek and its affiliates;
• third-party products and services “developed using DeepSeek implementations”, a wording quoted by The Register and BleepingComputer.

The duties for agencies, as summarised by BleepingComputer:

1. immediately exclude covered products from procurement, including approved-product lists and purchasing marketplaces;
2. immediately remove them from IT systems, or request an exception;
3. submit exception requests within 60 days of the order, with the justification and proposed mitigations; CISA reviews them and replies in writing. CyberScoop speaks of a “compelling justification”, such as the inability to perform a mission-critical function.

CISA will provide technical assistance to agencies that lack the capabilities to comply. The order is not a general ban: it does not apply to private or consumer use.

The risks cited by DHS

According to the letter quoted by CyberScoop, DeepSeek’s models present “multiple threat vectors” to agency systems:

• data stored on servers in China, where national security laws allow the government to request data from companies;
• collection of personal information and tracking of users’ “digital fingerprints” in models hosted directly by DeepSeek;
• the prompt and response mechanism, which may transfer sensitive information; combined with public data, it could produce intelligence on people and government programmes;
• possible malware or vulnerabilities in the models. BleepingComputer reports that the order cites a February 2025 NowSecure analysis of the DeepSeek iOS app: unencrypted transmissions, weak encryption algorithms, insecure local storage, user fingerprinting risks and data sharing with ByteDance.

Until full removal, the order encourages agencies to monitor inbound and outbound traffic, increase logging and data exfiltration controls on the affected systems, and flag anomalous activity.

The criticism: a scope that is hard to verify

The Register considers the order too vague. It is unclear what “developed using” means: ongoing access to the models via API, or any software written with DeepSeek’s help. Katie Moussouris, founder of Luta Security, notes that banning DeepSeek’s app and API is workable, but that an AI software bill of materials does not record which tools developers used: every vendor would have to attest that nothing it relies on ever touched DeepSeek, with no practical way to verify it. Moussouris also argues that the order goes too far by including US-hosted versions of the models, since the risks it cites are mostly about data reaching servers in China. She also mentions NIST evaluations finding DeepSeek more vulnerable to prompt injection and easier to jailbreak than US models.

The most practical point for infrastructure teams concerns AI gateways. Services such as OpenRouter and Vercel AI Gateway also offer DeepSeek’s models via API and, by default, choose the provider themselves: an organisation may therefore use DeepSeek without knowing it, unless it has restricted the allowed models and providers.

Background

According to The Register, DISA had been blocking DeepSeek on Pentagon networks since January 2025, and similar steps were taken by the US Navy, NASA, the House of Representatives and several states. CyberScoop points out that DeepSeek is the only Chinese model named in the order, but not the only one raising concerns: it also mentions Moonshot AI’s Kimi and Alibaba’s Qwen.

What it means for European and Italian companies

The order only applies to US federal civilian agencies, but the problem it describes affects everyone: knowing which AI models process company data, and where. For NIS2 entities, AI services fall under supply chain security and asset management (Article 21); for healthcare and anyone handling personal data, GDPR also matters, because every prompt may contain data that ends up with a non-EU provider. From a detection standpoint, data sent to an unauthorised AI service follows the pattern of exfiltration over web services (MITRE ATT&CK T1567).

What to do now

1. Map AI usage: apps, browser extensions, APIs called from code, coding agents on developers’ workstations.
2. Check gateways and libraries that route to multiple providers: set an explicit list of allowed models and providers and turn off automatic selection.
3. Ask your vendors which models their products use, where they are hosted and whether prompts are retained or used for training.
4. Watch outbound traffic: DNS and proxy logs towards AI service endpoints, with alerts for unauthorised destinations and abnormal volumes.
5. Write a policy stating which data may go into prompts and which services are allowed.

How ARXDOME helps

The ARX Sentry agent, on Windows, Linux and macOS, collects the inventory of software, processes and ports on each machine: it helps find AI apps and tools installed without authorisation. The NOC map shows outbound traffic blocked by the firewall and connections to sources on threat lists, and the SOC pages bring together software, events and MITRE ATT&CK tactics for each machine.

Last updated: 20 October 2026