AI supply chain security: checking model provenance before you deploy
Build an approval process for models, datasets and AI libraries so you know where every component came from and whether it has been tampered with.
This is for CTOs, platform leads and security heads whose teams download open models, embedding models, datasets and AI libraries to build features. Engineers often pull a promising model from a public hub and ship it within days. That component may have an unclear origin, a restrictive licence, or a file format that can execute code when loaded. This post gives you a lightweight approval process so every AI component has a known source, verified integrity and a named owner.
A data science team finds an embedding model on a public hub that performs well on Hindi and English text. They download it from a personal account with few details, load it in a notebook, and move it into production. Nobody checks who published it, whether the file was modified, or whether its format can run arbitrary code during loading.
Six months later, a customer's procurement team asks for a list of AI components, their sources and licences. Nobody has one. Different teams use different copies of similar models, a few with licences that restrict commercial use. When a library used for model serving announces a security fix, the team cannot quickly tell which services depend on it.
A healthcare software company in Hyderabad used several open models for document classification and medical text extraction, each downloaded by different engineers over a year. A hospital client's IT team asked for a component list with sources and licences during contract renewal. The CTO set up an intake form, had security and legal review every model in use, replaced two with clearer licensing, converted weights to a tensor-only format, and moved all approved files into an internal registry with pinned versions. After the change, the team answered the client's questionnaire with a clear AI bill of materials, new models went through a short review, and a later library security advisory was mapped to affected services within the same day.
What is an AI bill of materials?
It is an inventory of AI components in your products: models, datasets, embedding models and key libraries, with source, version, licence, owner and the services using each. It lets you answer customer questions and respond quickly to advisories.
Are hosted model APIs part of the supply chain too?
Yes. Record the provider, model version, data handling terms and region for each hosted API you call. Track version changes, because a provider update can change behaviour even when your code does not.
Why do model file formats matter?
Some older formats can execute code when a model is loaded, so a tampered file could run commands on your server. Tensor-only formats store weights without executable content, which removes that risk during loading.
Who should approve a new model?
The engineering owner confirms technical fit, security checks source and integrity, and legal confirms the licence allows your use. For models touching personal or health data, include your data protection lead as well.
Want this fixed in your business?
Send the workflow. You get a scope and a fixed price within two working days — or a straight answer that it isn't a fit.