What happened
DHS Exclusion and Removal Order 2026-01 has an original release date of July 24, 2026 and was published in the Federal Register on October 20, 2026. Order 2026-01 applies to Federal Civilian Executive Branch agencies except national security systems and secure compartmented information systems. The order requires agencies to exclude DeepSeek and affiliates from civilian procurements and remove DeepSeek, affiliate, and third-party products or services developed using DeepSeek implementations. The public notice makes the downstream scope explicit: products sold under another provider’s name may still be covered when their service was developed using a DeepSeek implementation.
Exception requests had to be submitted in writing within 60 days of issuance and include compelling justification and alternative mitigations. The order encourages traffic monitoring, enhanced logging, data-exfiltration controls and anomalous-activity detection until covered products are fully removed. DHS cites data location, legal access, personal-information collection and potential malware or security vulnerabilities as risk grounds; the order does not report a confirmed agency compromise. Attribution posture: The DHS order attributes no cyber incident or actor; it records a government supply-chain risk decision concerning DeepSeek.
The cited order concerns DeepSeek products and implementations; it does not identify a specific agent or framework. No underlying DeepSeek model or version is identified in the cited order. The order covers direct DeepSeek offerings and third-party products and services developed using DeepSeek implementations; it does not describe operator configurations. The policy action requires covered agencies to exclude affected procurements and remove covered implementations; no AI-executed mechanical action is alleged.
Why this matters now
The operative scope reaches beyond a named chatbot, website or API. It includes products and services from DeepSeek and its affiliates, plus third-party products and services developed using DeepSeek implementations. That creates a discovery problem across SaaS features, model gateways, development tools, embedded assistants and subcontractor deliverables. Network blocks cannot prove compliance when the underlying model is abstracted behind another provider’s interface.
The order is directly binding on covered federal civilian agencies, not a general prohibition for private companies. Its wider enterprise significance is the evidence standard it introduces: buyers need product-level model provenance, suppliers need defensible attestations, and exceptions require documented mission impact and alternative mitigations. Organisations selling into the federal market should expect security, procurement and legal teams to reconcile contract scope with technical inventories rather than rely on vendor brand names.
The decision for security leaders
Federal CISOs should pair technical discovery with procurement and contract records. Require each system owner and supplier to identify the underlying model provider, hosting path and third-party implementation. Where provenance cannot be established, classify the product as unresolved rather than compliant. Validate removal through configuration evidence, network observations and supplier attestations, not licence cancellation alone.
Private-sector leaders should not misstate the order as a universal legal ban. Instead, use it to test whether AI inventories can answer a practical provenance question across SaaS and supply chains. Federal suppliers should involve legal counsel and contracting officers where deliverables, subcontractors or managed services may contain covered implementations. Exceptions should be explicit, justified, mitigated and time-bound.
Evidence of closure
- An approved inventory identifies model provider and implementation for each AI-enabled system.
- Supplier attestations address products, services and subcontractor-delivered capabilities.
- Technical validation shows covered implementations are removed or access is disabled.
- Each exception records justification, alternative mitigations and accountable approval.
The Security.io assessment
The newly visible decision is not simply whether employees can access a DeepSeek website. It is whether organisations can prove which model implementations sit inside products acquired under other names. That is an AI supply-chain governance test, and many conventional software inventories do not record model provider, model version, hosting route or prompt-data path with sufficient precision.
The order documents a government risk judgment; it does not establish that DeepSeek compromised an agency, that every DeepSeek implementation behaves identically, or that private enterprises have the same legal duty. Security leaders should preserve that distinction while strengthening provenance and exception controls. Our assessment changes if CISA narrows or expands implementation guidance, agencies publish exception decisions, or suppliers disclose material embedded use.
Questions for the morning meeting
- Can the organisation identify underlying models used inside every AI-enabled product and service?
- Which suppliers can attest that federal deliverables contain no covered DeepSeek implementation?
- Who approves and time-bounds exceptions where removal affects mission-critical services?