Cardboard Iguana Security

Properties6
created2026-10-21
updated2026-10-21
typeconcept
confidencemedium
sourceshttps://www.federalregister.gov/documents/full_text/text/2026/10/20/2026-21447.txt, https://www.bankinfosecurity.com/dhs-bans-deepseek-from-federal-civilian-agencies-a-33146, https://thehackernews.com/2026/10/us-orders-deepseek-removed-from-federal.html
tagssecurity, political-science, ai, reference

8 min read

Government Supply-Chain Bans on Adversary Software

Governments have two levers against software they consider a national-security risk: ban it on their own systems, or ban everyone from selling it. The United States’ sharper instrument is the exclusion and removal order under the Federal Acquisition Supply Chain Security Act of 2018 (FASCSA) — a directed exclusion from procurement plus an affirmative order to rip running products out of agency networks. FASCSA reached its highest-profile target in October 2026, when DHS ordered DeepSeek products out of every federal civilian agency, including third-party products “developed using DeepSeek implementations.” The derivative clause is what makes the case durable reading: it extends a vendor ban into the open-weights ecosystem, where provenance is exactly what nobody can see.

The instrument: FASCSA exclusion and removal orders

FASCSA (41 U.S.C. 1321–1328, implemented at 41 CFR 201-1) puts the Federal Acquisition Security Council (FASC) in the recommending role and leaves issuance to three officials: the Secretary of Homeland Security for civilian agencies, the Secretary of Defense for the Defense Department and national security systems, and the Director of National Intelligence for the intelligence community. Executive agencies “shall comply with exclusion and removal orders”; an agency wanting relief must ask the issuing official for a written exception with compelling justification, administered for DHS orders by CISA’s FASCSA Order Implementation Team. The issuing official must notify the named company and the relevant congressional committees and must review the order at least once a year. A targeted company can challenge the order in the U.S. Court of Appeals for the D.C. Circuit within 60 days of receiving notice. 12

Orders reach past agencies into their contractors. FAR clause 52.204-30 bars a contractor from providing or using any product or service covered by a FASCSA order while performing the contract, requires a SAM.gov search for “FASCSA order” at least quarterly, reasonable supply-chain inquiry when a new order lands, reporting of any covered product within 3 business days, mitigation information within 10, and flow-down into subcontracts. The FASC’s implementing rule extends orders to prime contractors and subcontractors, and a 2023 FAR update makes contractor compliance mandatory. 34

The law was first used in 2025, when the DNI excluded Acronis, a Swiss data-protection vendor, from intelligence-community procurement and ordered its products removed from IC systems. 5

The DeepSeek order (2026)

Exclusion and Removal Order 2026-01, signed by DHS Secretary Markwayne Mullin on July 24, 2026 after a FASC recommendation, does two things: it excludes IT products and services developed by Hangzhou DeepSeek Artificial Intelligence Co., Ltd and its affiliates from federal civilian agency procurements, and it orders removal from civilian agency IT of all DeepSeek and DeepSeek-affiliate products plus “third-party products and services developed using DeepSeek implementations.” DHS filed the order for public inspection on October 19 and published it in the Federal Register on October 20 (91 FR 66346, docket DHS-2026-0628); the signed order itself is four scanned pages attached to the notice. 67

The Federal Register notice gives no reasons. Reporting on the order and the DHS position cites “multiple threat vectors”: use of the models “may result in the transfer of sensitive data to the PRC” by employees or vendors, Chinese law lets state cyber and intelligence services reach data on PRC-domestic infrastructure, the models “may contain malware or other security vulnerabilities” (citing research on critical flaws in DeepSeek’s iOS app), and agency use “could produce actionable intelligence on U.S. persons and/or government agencies.” The order also urges agencies to monitor network traffic for data leaving their environments and to tighten logging. 89

Two procedural facts stand out. The order is dated July 24 but was filed publicly 87 days later, and the FASC regulations set no deadline for public release — so the agency exception window (60 days from issuance, by the terms of the order itself) largely ran before agencies could see the document. And the order binds what civilian agencies buy and run; it does not restrict companies or the public, and it excludes the Department of Defense. 10

The derivative-reach problem

The clause worth watching is “developed using DeepSeek implementations.” DeepSeek ships MIT-licensed open weights (see deepseek), so a ban on the vendor’s products normally evaporates at the first fine-tune; this clause attempts to follow the model into derivatives — tools built on DeepSeek implementations, which reporting reads as likely including products built on its open-weight models. The notice does not define “DeepSeek implementations,” and enforcement means establishing provenance for a model artifact (fine-tunes, distillations, quantizations) that open release deliberately makes untraceable. The House Select Committee’s April 2025 “DeepSeek Unmasked” report alleged DeepSeek itself leaned on distillation of U.S. models — provenance questions cutting both directions. 1112

This is the structural tension with gongkai: open release is a deliberate ecosystem strategy (give away the artifact, own the position), and supply-chain exclusion is the state’s attempt to keep jurisdiction over artifacts whose whole point is redistribution. Any watermarked or attested-model scheme (ai-text-watermarking) is doing the same work from the technical side.

The wider pattern

Before the order, U.S. restrictions on DeepSeek were device bans adopted one agency at a time: by February 2025 Congress, NASA, the Navy, the Pentagon, and Texas had barred it from government devices, with New York and Virginia following within weeks; Australia, South Korea, and Taiwan keep it off government systems. Federal bills to bar DeepSeek on government devices were introduced in both chambers in February 2025 and had not advanced past committee by October 2026. 1314

The same statute has been aimed at an American AI lab: according to BankInfoSecurity, the Department of Defense invoked FASCSA to blacklist Anthropic, and Anthropic sought full appeals-court review of the Pentagon ban. The proposed No Adversarial AI Act would generalize the DeepSeek approach into a standing ban on foreign-adversary AI models in federal IT. China’s foreign ministry responded to the DeepSeek order on October 20 (spokesperson Mao Ning, via AP): Beijing “opposes the overstretching of the concept of national security” and will “resolutely safeguard the legitimate rights and interests of Chinese companies.” 1516

Open questions

How will “developed using DeepSeek implementations” be adjudicated when a vendor disputes it — is a fine-tune covered, is a distilled student model covered, and who proves what? Agencies that need an exception after the window closed have no obvious path. The annual-review requirement gives the order a built-in sunset check without a sunset. And the Acronis, DeepSeek, and Anthropic cases together leave untested where the D.C. Circuit draws the line on evidence: DeepSeek’s challenge right expires 60 days after notice, and the notice itself states no reasons.

  • deepseek — the entity the order targets; open weights and a China-siting data position that the order’s threat model draws on
  • gongkai — open release as ecosystem strategy, and the state’s counter-move examined here
  • china-ai-datacenter-buildout — the compute layer the export-control regime is already constraining; software exclusion is the same policy family pointed at models
  • model-backdoors — the supply-chain suspicion underneath “may contain malware or other security vulnerabilities”; hardware analog in ledger-nanox-spy-implant-2026
  • ai-text-watermarking — the technical side of artifact provenance, the problem the derivative clause runs into

Sources

Footnotes

  1. Swati Khandelwal 2026 — U.S. Orders DeepSeek Removed From Federal Civilian Systems, Bans New Purchases ↩

  2. Chris Riotta 2026 — DHS Bans DeepSeek From Federal Civilian Agencies ↩

  3. Swati Khandelwal 2026 — U.S. Orders DeepSeek Removed From Federal Civilian Systems, Bans New Purchases ↩

  4. Chris Riotta 2026 — DHS Bans DeepSeek From Federal Civilian Agencies ↩

  5. Swati Khandelwal 2026 — U.S. Orders DeepSeek Removed From Federal Civilian Systems, Bans New Purchases ↩

  6. U.S. Department of Homeland Security 2026 — Notice of Order Regarding Exclusion and Removal of DeepSeek Products From Federal Civilian Executive Branch Information Technology Systems and Procurements ↩

  7. Swati Khandelwal 2026 — U.S. Orders DeepSeek Removed From Federal Civilian Systems, Bans New Purchases ↩

  8. Chris Riotta 2026 — DHS Bans DeepSeek From Federal Civilian Agencies ↩

  9. U.S. Department of Homeland Security 2026 — Notice of Order Regarding Exclusion and Removal of DeepSeek Products From Federal Civilian Executive Branch Information Technology Systems and Procurements ↩

  10. Swati Khandelwal 2026 — U.S. Orders DeepSeek Removed From Federal Civilian Systems, Bans New Purchases ↩

  11. Swati Khandelwal 2026 — U.S. Orders DeepSeek Removed From Federal Civilian Systems, Bans New Purchases ↩

  12. Chris Riotta 2026 — DHS Bans DeepSeek From Federal Civilian Agencies ↩

  13. Swati Khandelwal 2026 — U.S. Orders DeepSeek Removed From Federal Civilian Systems, Bans New Purchases ↩

  14. Chris Riotta 2026 — DHS Bans DeepSeek From Federal Civilian Agencies ↩

  15. Chris Riotta 2026 — DHS Bans DeepSeek From Federal Civilian Agencies ↩

  16. Swati Khandelwal 2026 — U.S. Orders DeepSeek Removed From Federal Civilian Systems, Bans New Purchases ↩

Backlinks

Recent notes

See 577 more →

Graph View