Agent Upstart

Cool things people are building with agents — and practical help getting the most out of them.

Playbooks · Teams

DHS DeepSeek order: inventory your stack

Written November 7, 2026

DHS ordered civilian agencies to remove DeepSeek — including third-party products built on it. If you sell to them, your dependency inventory is the test.

What the order says

The Department of Homeland Security issued Exclusion and Removal Order #2026-01 under the Federal Acquisition Supply Chain Security Act of 2018. The order was signed by Secretary Markwayne Mullin on July 24, 2026, and published as a Federal Register notice on October 20, 2026. It does two things, both effective immediately: it excludes information technology products and services developed by Hangzhou DeepSeek Artificial Intelligence Co., Ltd., including affiliated entities, from federal civilian agency procurements; and it requires the removal of all DeepSeek and DeepSeek-affiliate products, plus third-party products and services developed using DeepSeek implementations, from civilian agency information technology.

The scope is federal civilian executive branch agencies, except for their national security systems and secure compartmented information systems. It is the first FASCSA order issued by DHS, and the first to cover all federal civilian executive branch agencies. Two earlier FASCSA actions were narrower: the Office of the Director of National Intelligence issued the first, against Acronis, in September 2025, and the Department of War designated Anthropic in March 2026. The order is not a general prohibition on private companies or the public. It covers federal civilian systems and purchasing only.

Agencies could request an exception in writing within 60 days of issuance, with a compelling justification, under the process in 41 CFR 201-1.304(b). Counted from the July 24 signing, that 60-day window closed before the order appeared in the Federal Register, a timing problem the Cloud Security Alliance analysis flags directly. Contractors have a separate route: where the FAR 52.204-30 clause is in a contract, they may disclose and request a waiver when the order is added to the contract by modification.

The phrase that does the work

The removal clause reaches "third-party products and services developed using DeepSeek implementations." The order never defines "DeepSeek implementations," and that gap is the whole inventory problem. The Cloud Security Alliance sets out how the phrase could plausibly reach, in decreasing order of certainty:

That list is an interpretation for risk planning, not a ruling. The Alliance says so, and notes that CISA had published no implementing guidance on the "developed using" standard when the order became public. The order's stated rationale is about DeepSeek's servers being in China, prompts that could transfer sensitive data there, and the possibility that the models contain malware or other vulnerabilities. The removal language is broader than that rationale, and it is the broader language a vendor has to answer to.

The inventory

If your product reaches a federal civilian agency, build the inventory before anyone asks for it. Work through these, and write down the answer for each, with a date:

  1. Runtime dependencies. Every model your product calls at runtime, named by base model and host: the DeepSeek app or API, any self-hosted DeepSeek weights, any model served through a gateway or cloud account. Gateways, cloud accounts, managed applications, and approved local runtimes each need a line showing no permitted or observed DeepSeek route.
  2. Weights and lineage. Every weight file in the product, with its base model recorded. Fine-tunes, adapters, quantizations, and merges need their base model named, since that is the field that reveals a DeepSeek parent.
  3. Training and evaluation inputs. Whether DeepSeek outputs were used for distillation, synthetic training data, or evaluation. Few organizations keep records that can show this, which is exactly why the answer is usually "unknown" rather than "no." Record it as unknown where it is unknown; a blank is not an answer.
  4. Development tooling. Whether any coding assistant or agent in the build pipeline runs on a DeepSeek-backed model, and whether that tooling is logged. This is the hardest row to answer honestly, and the one an agency customer is most likely to ask about next.
  5. Subcontractors and suppliers. Where FAR 52.204-30 is in a contract, it requires contractors to search SAM.gov for applicable FASCSA orders, report covered articles they find within three business days, and flow the clause down to subcontractors. Ask the same inventory questions of every supplier whose model or code ships inside your product.

Keep the inventory as an AI bill of materials: base models, fine-tune lineage, distillation sources, and hosted endpoints, per product, per version. Runtime use is detectable today with the tooling that exists. Development-time use usually is not recorded at all, and no tool can reconstruct it from weights after the fact. That asymmetry is the point of writing the inventory down: the parts you can prove, the parts you cannot, and the date each was checked.

Limits

Related playbooks

Sources

Back to: Teams playbooks hub · Also: Audit apps that installed agent plugins for you · Audit a repository's git config before any agent opens it · Check your Codex and Claude Code config for relay tampering