The Department of Homeland Security (DHS) has mandated that federal civilian agencies remove DeepSeek products and block them from contracts. This order affects IT products and services from Hangzhou DeepSeek Artificial Intelligence Co. and its affiliates. It also extends to third-party products "developed using DeepSeek implementations," likely including tools built on open-weight models. This second clause presents a significant challenge for most teams.
This checklist outlines what procurement-security teams should verify now. It applies to agencies and any contractor or subcontractor serving them. Each item has a clear done/not-done state and a description of what "good" looks like.
Prerequisites
Before you begin, ensure you have:
- A comprehensive inventory of software, SaaS, and AI-enabled tools in use, with named owners.
- Access to network logging and proxy or DNS telemetry.
- Contract repositories that your team can search.
- A designated decision-maker for exception requests and a legal contact to interpret the order.
If you're missing any of these, addressing that gap is your first task.
Checklist
1. Confirm your legal exposure.
Complete this step by documenting whether the order applies to you directly (as a civilian agency) or indirectly (as a prime or subcontractor). The order was issued under the Federal Acquisition Supply Chain Security Act of 2018. The Federal Acquisition Security Council's implementing rule states that exclusion and removal orders can reach prime contractors and subcontractors serving federal agencies. A 2023 Federal Acquisition Regulation update requires contractors to follow these orders.
Good looks like: A one-page legal determination, signed by counsel, filed with your compliance records.
2. Inventory direct DeepSeek use, including shadow use.
Complete this step by searching endpoints, browser extensions, API keys, developer repositories, and expense reports for DeepSeek apps and endpoints. Several agencies had already blocked the chatbot before this order, so employee-level use is a known pattern.
Good looks like: A dated inventory with each instance either removed or tied to a pending exception request.
3. Trace embedded model dependencies in third-party tools.
Complete this step by obtaining written confirmation from every vendor with AI features about which foundation models power their product. This is the most challenging control. The removal mandate affects tools built on DeepSeek's models, so a vendor you've never associated with DeepSeek may still be in scope. Treat it as a Fourth and Nth Party Management problem, not a one-time questionnaire.
Good looks like: A model-lineage record for each in-scope vendor, including the vendor's attestation and the date it was obtained.
4. Update Vendor Intake to screen for model provenance.
Complete this step by ensuring your intake workflow requires every AI-enabled vendor to disclose underlying models, hosting locations, and Sub-Processor Disclosure details before a contract is signed. This should be part of your Pre-Contractual Assessment, not your annual review.
Good looks like: Intake cannot be completed without a provenance answer, and "unknown" triggers escalation.
5. Amend contracts with a Sub-Outsourcing Clause covering models.
Complete this step by requiring new and renewing agreements to notify you before changing or adding a foundation model, and to warrant compliance with exclusion and removal orders. Add Termination Rights tied to a government exclusion order, and a Right to Audit that reaches model sourcing.
Good looks like: Standard clause language approved by legal and applied to every AI-related renewal.
6. Monitor network traffic for data leaving your environment.
Complete this step by tuning detection to flag traffic to DeepSeek-associated domains and unexplained outbound transfers. The order urges agencies to monitor network traffic for signs of data leaving their environments and to tighten logging. DHS specifically pointed to the risk of sensitive data moving to China and to DeepSeek's China-based servers.
Good looks like: Alerts routed to a named responder, log retention confirmed, and a documented test showing the detection fires.
7. File exception requests inside the 60-day window.
Complete this step by submitting any justified exception within 60 days of the order's issuance. The Cybersecurity and Infrastructure Security Agency approves or denies requests in writing. Don't assume silence means approval.
Good looks like: Each request carries a business justification, a compensating-controls description, and a calendar reminder for the deadline.
8. Build a Wind-Down Plan for affected tools.
Complete this step by ensuring each removed or replaced tool has a documented migration path, Access Revocation steps, and a data-deletion confirmation from the vendor. Assess Substitutability before you pull a tool that supports a Critical or Important Function.
Good looks like: Tested cutover steps and written vendor confirmation that your data has been deleted.
Common Mistakes
Scoping to the brand name only. Searching for "DeepSeek" misses vendors that wrap its open-weight models under a different product name. Scope to the models, not the logo.
Relying on a single questionnaire. A SIG Core response won't tell you which model runs behind an AI feature unless you ask. Add targeted questions, and verify answers against contract terms and observed network behavior.
Treating Cyber Risk Ratings as proof. A Cyber Risk Rating measures external security hygiene. It says nothing about model provenance or legal exposure to foreign government access.
Missing the clock. The 60-day exception window is short. Teams that start with inventory in week eight will miss it.
Ignoring the geopolitical pattern. Texas, New York, and Virginia restricted DeepSeek on state devices, and Australia, South Korea, and Taiwan keep it off government systems. Multiple jurisdictions are acting, so your Risk Governance Framework should expect more country- and model-level restrictions, not treat this as a one-off.
Next Steps
Start with items 1 and 3 this week; they set your scope and expose your largest blind spot. Then integrate model-provenance questions into Vendor Intake permanently, so the next order doesn't trigger a scramble.
Incorporate AI model lineage into your Continuous Monitoring of Active Arrangements. Vendors often change underlying models without much notice, so a clean answer today needs a contractual duty to update it. Finally, conduct a Concentration Risk Assessment on your AI providers. If one foundation model underpins several of your tools, a single government order can disrupt all of them at once. You'll want to know that before an order arrives.
