Explore our new Cybersecurity Terms Glossary, explained by our experts.
Cybersecurity Terms Glossary
AI supply chain security addresses everything an AI system inherits from elsewhere: pre-trained weights, datasets, adapters, frameworks, and hosted APIs, any of which can arrive compromised.

What Is AI Supply Chain Security?
Almost nobody trains a model from scratch. A typical AI feature is a downloaded set of weights, a fine-tuning adapter, a handful of datasets, several Python frameworks, a vector store, and a hosted API—most of it pulled from the internet by a data scientist on a deadline. AI supply chain security is the work of knowing what those components are, where they came from, and what they can do once they are inside your environment.
The most under-appreciated fact in this area is that many model formats are not inert data. PyTorch checkpoints have historically been Python pickle files, and loading one can execute arbitrary code; Keras models could carry a Lambda layer containing code, addressed by CVE-2024-3660. In February 2024, JFrog reported roughly a hundred models on Hugging Face containing malicious payloads, several of which opened reverse shells when loaded. The practical response is to prefer the safetensors format, which stores tensors without code, scan model files before loading, and treat any load of an untrusted checkpoint as code execution in your environment.
You cannot govern components you have not listed. Extend the software bill of materials approach to AI assets: record every model with its source, version, licence, and hash; every dataset with its provenance and the lawful basis for its use; every adapter and the base model it modifies; and every external AI service with what data it receives. CycloneDX added a machine learning bill of materials profile for exactly this, and the EU AI Act’s documentation duties for general-purpose models push in the same direction. A register also answers the question that follows every disclosure—do we use this model, and where.
Pull models through an internal mirror rather than directly from a public hub, so you control versions and can block a withdrawn artefact. Pin and verify hashes, because a model repository can be updated in place. Load untrusted weights first in an isolated environment with no credentials and no internal network route. Require the safetensors format where the ecosystem supports it. Apply the same dependency discipline to ML libraries as to the rest of your estate, including an internal package proxy that fails closed on unknown names—the control that also defeats hallucinated dependencies described in our note on AI-generated code. Finally, evaluate behaviour after import: a backdoor that triggers on a specific phrase will not show up in accuracy metrics, so include targeted probing in your AI red teaming.
Supply chain security is already a named expectation in NCSC guidance, both in the Guidelines for Secure AI System Development and in the supply chain guidance that underpins Cyber Essentials and the Cyber Assessment Framework. For organisations in scope of the NIS Regulations, an AI component inside an essential service is part of the supply chain you must manage. The EU AI Act adds documentation and transparency duties for providers of general-purpose models, which in turn gives downstream deployers something to ask for. At Privia Security, we review model provenance, ML infrastructure, and connector permissions as part of our Infrastructure and Architecture Review Services.
It is routine, but treat it as importing executable code. Prefer well-known publishers, verify the hash against the version you reviewed, choose safetensors over pickle-based formats, and load anything unfamiliar in an isolated environment with no credentials. Mirroring models internally gives you version control and a single place to block an artefact that turns out to be malicious.
It is an inventory of the models, datasets, adapters, and frameworks that make up an AI system, with source, version, hash, and licence for each. CycloneDX supports it as a machine learning profile. The value is operational: when a model or library is found to be compromised, you can answer within minutes whether you use it and which services would be affected.
Only partly. A backdoor triggered by a rare phrase or pattern will not change aggregate accuracy, so standard evaluation misses it. Targeted probing, behavioural comparison against a reference model, and checking outputs on sensitive topics all help, but provenance remains the stronger control. Use models you can trace to a publisher and a verified artefact.
Yes. An adapter is small and easy to share, which is exactly why it circulates without review, and it can change behaviour significantly, including weakening safety training built into the base model. Record which base model each adapter modifies, store adapters in your own repository, and re-evaluate behaviour after applying one rather than assuming the base model’s evaluation still holds.
You May Be Interested In These
We use a few cookies that are needed to make this website work, and they are always on. We would also like to set analytics cookies, to see which pages are useful; functional cookies, so that our live chat works; and advertising cookies, to measure our campaigns. We will not set any of those until you say yes, and you can change your mind at any time from the Cookie Settings link in the footer.
Choose the cookies you are happy for us to use. Strictly necessary cookies are always on; everything else is off until you turn it on.
These cookies keep the website working: they hold your security checks, your form submissions and the choice you make here. The site cannot run without them, so they are always on.
These let our live chat open and remember your conversation, and let the newsletter sign-up work. Switch them off and the chat window will not appear.
These tell us which pages are read and how people find us, so that we can improve the site. The reports are about groups of visitors, never about you personally.
These let us measure our advertising and show you our adverts on other websites, such as LinkedIn, Google and Meta. Switch them off and you will still see adverts, but they will be less relevant.