Audit the AI Models Hidden Inside Vendor Software
Matthew Leo · Published October 23, 2026 · Guides
A new U.S. government order does more than remove DeepSeek's own applications from federal civilian systems. It also reaches third-party products and services developed using DeepSeek implementations.
That wording turns an app-ban story into a software-supply-chain problem. A buyer may know the name on the contract without knowing which model powers a feature, where prompts are processed or whether the vendor can switch providers without notice.
The Department of Homeland Security order applies to U.S. federal civilian executive-branch information systems. It is not Canadian law. Reuters and The Associated Press independently reported the order and its security rationale.
Why a product name is not enough
AI features are often assembled from several layers: an application vendor, a model provider, a cloud host, a retrieval service and monitoring or safety tools. A vendor may also keep a fallback model for capacity or price reasons.
That means a procurement record that says only “uses generative AI” cannot answer basic questions during a policy change or security incident. Teams need a component-level record that can be updated.
The model-provenance inventory
For every AI-enabled feature, record the following in plain language:
- Product and feature: the exact function that invokes a model, not just the vendor's product family.
- Model provider and model family: include the version when the vendor exposes it.
- Deployment path: vendor-hosted, customer-hosted or sent to a separate application-programming interface.
- Data path: what leaves your environment, where it is processed, how long it is retained and whether it is used for training or evaluation.
- Subprocessors: the companies that can receive prompts, files, embeddings, logs or output.
- Fallbacks and routing: whether the vendor can send requests to a different model when the primary service is unavailable.
- Change control: how much notice you receive before the model, host or subprocessor changes.
- Exit path: how to disable the feature, select another model or export the data needed to move.
Questions to put in the contract
Ask the vendor to warrant that its model and subprocessor list is complete. Require notice before a material model, hosting-jurisdiction or data-use change. Set a response time for security questions and a deadline for removing a prohibited component.
Also ask what evidence the vendor can provide. Useful evidence may include current architecture diagrams, subprocessor lists, data-flow documentation, independent assurance reports and records showing that a disabled AI feature no longer sends data.
Do not accept “we do not train on your data” as a complete answer. Retention, human review, abuse monitoring and subprocessors can still matter. Mapletechie's guide to AI data-retention promises explains the questions hidden behind that phrase.
Use risk tiers instead of one rule for everything
A writing assistant used on public marketing copy does not require the same review as a system handling health, legal, financial or security data. Put higher-risk uses through deeper technical and legal checks, and restrict sensitive inputs until the evidence is adequate.
The Government of Canada's guide on the use of generative AI emphasizes assessing security, privacy and legal risks before deployment. Innovation, Science and Economic Development Canada's implementation guide for managers likewise calls for procurement due diligence and vendor transparency.
Run a substitution drill
Choose one important AI feature and simulate the loss of its current model provider. Can the feature be disabled without breaking the rest of the product? Can stored prompts and outputs be deleted? Is there a reviewed replacement, and would switching it change accuracy, privacy or cost?
A model inventory is useful only if it supports a decision. The U.S. DeepSeek order shows how quickly a named component can become unacceptable to a buyer. Organizations that can trace model dependencies can respond with evidence instead of guessing which products are affected.
Sources: U.S. Department of Homeland Security order; Reuters; Associated Press; Government of Canada; ISED.
Tags: AI procurement, DeepSeek, Vendor risk, Model provenance