DHS Orders Federal Agencies to Remove DeepSeek, Including Third-Party Products Built on It. Do You Know Where It Is in Your Stack?

The US Department of Homeland Security has ordered federal civilian agencies to stop using DeepSeek products. According to BleepingComputer, the "notice of order" was published in the Federal Register on 20 October 2026 (document 2026-21447). It covers DeepSeek's own products, those of its affiliates, and third-party products and services built using DeepSeek technology. It is not a general ban on private or consumer use.

What the order requires

• Procurement: covered products must be excluded immediately, including from approved-product lists and purchasing marketplaces.
• Removal: agencies must remove covered products from their IT systems immediately, or request an exception.
• Exceptions: requests are due within 60 days of the order, with a justification and proposed risk mitigations. CISA reviews each one and issues a written approval or denial. Agencies that cannot comply because of internal capability limits can get technical assistance from CISA.

Why DHS says it did this

The order cites concern that the Chinese government could access DeepSeek infrastructure, collect personal information, track users, or receive sensitive government data. It also raises the possibility of malware in DeepSeek models, pointing to a February 2025 NowSecure report on the DeepSeek iOS app that found unencrypted data transmission, weak encryption, insecure local storage, and data sharing with ByteDance.

The order does not say any federal system was actually compromised. It is a risk-based exclusion, not an incident response.

The hard part is the third-party clause

Removing an app someone installed is easy. The phrase "third-party products and services built using DeepSeek technology" is not. A model can sit behind a vendor's summarisation feature, a coding assistant's backend, a support chatbot, or an open-weights model fine-tuned and repackaged under another name. Procurement paperwork rarely says which model is underneath.

That is a problem well beyond government. Customers, auditors and insurers tend to copy federal restrictions, and similar questions will show up in vendor questionnaires: which models do your staff and your suppliers send data to, and can you prove none of them are on a restricted list? We covered a related exposure in the DeepSeek Harness sandbox escape, where the risk came from the tool wrapped around a model rather than the model itself.

What to do this week

1. Ask directly. Survey teams for any use of DeepSeek apps, APIs or self-hosted weights, including personal accounts used for work.
2. Check network and SaaS logs for DeepSeek domains and API endpoints, and for browser extensions or desktop tools that call them.
3. Ask your AI vendors in writing which underlying models process your data, and whether any are DeepSeek or derived from it. Record the answer.
4. Write an approved-model list and decide what happens to a request for a model that is not on it.
5. Decide your policy before a customer asks. Even without a legal obligation, you will want a documented position.

The question to take away

If a regulator, customer or insurer asked tomorrow which AI models touch your data, could you answer from a list rather than a survey?

Talk to us about an approved-model list for your organisation.