Skip to content
StrikeCyberStrikeCyber
Field Manual · Attack technique

AI Supply Chain Security

Reviewed 13 November 2026Attack technique 2 min read

Quick answer

AI supply chain risk covers everything an AI application depends on but did not build: pre-trained models, fine-tuning adapters, datasets, AI libraries, model hosting providers and plugins. A malicious model file can run code when loaded, and a tampered model can behave normally until triggered. It is ranked third in the OWASP Top 10 for LLM Applications (2025).

On this page
01

How It Works

Most AI applications are assembled from parts. Teams download open-weight models and adapters from public hubs, pull datasets from the internet, install fast-moving libraries for orchestration and retrieval, and call hosted models through APIs. Each is a point of trust. Some model file formats, notably those built on Python's native object loading, can execute arbitrary code when opened, so simply loading an untrusted model can compromise the machine. Attackers also publish models under names that imitate popular ones, upload backdoored models that respond to a hidden trigger, and seed public datasets with poisoned data. The AI libraries themselves have had serious vulnerabilities, and Model Context Protocol servers and plugins add third-party code with access to tools and data.

02

Why It Matters

AI components often bypass the controls applied to ordinary software. Data scientists download models straight to GPU servers holding sensitive training data and cloud credentials, and model files are rarely scanned or recorded in software inventories. A backdoored model is especially hard to detect because it passes normal accuracy testing. MITRE ATLAS records AI supply chain compromise as a technique attackers use to gain initial access to AI systems.

03

How to Test for It

  1. Build an inventory of every model, adapter, dataset, AI library and hosted model provider in use, with versions and sources.
  2. Scan model files for formats that can execute code on load, and check whether any pipeline opens such files from untrusted sources.
  3. Verify provenance: are models pulled from the publisher's verified namespace, pinned to a specific version and checked against a known hash?
  4. Run dependency scanning on AI libraries and check for known exploited vulnerabilities in orchestration, serving and vector database components.
  5. Evaluate third-party models for unexpected behaviour on adversarial and trigger-style inputs before deployment.
  6. Review hosted model providers' terms on data retention, training on customer data and subprocessors.
04

How to Reduce the Risk

Prefer weight-only formats such as safetensors, and load anything untrusted in an isolated sandbox without credentials. Mirror approved models into an internal registry, pin versions and verify hashes. Maintain a software bill of materials that includes models and datasets, patch AI libraries as quickly as other internet-facing software, review the publishers of models and plugins, and red team any model before it reaches production. Apply the same vendor assessment to model providers as to any other supplier handling your data.

05

Regulatory and Framework Context

In Australia, ASD's ACSC co-authored international guidance on deploying AI systems securely and on AI data security, both of which address model provenance and supply chain integrity. AI suppliers also fall under third-party risk obligations such as APRA CPS 234 and CPS 230 for regulated entities, and ISO/IEC 42001 requires controls over third-party AI components.

Latest intel on this topic

Updated hourly from the daily briefing.

Common questions

Is it safe to download models from public model hubs?

It can be, with controls: use verified publishers, weight-only file formats, pinned versions and hash checks, and load new models in an isolated environment first.

How is this different from a normal software supply chain attack?

It includes everything in a normal software supply chain plus models and data, which cannot be reviewed like source code and can hide backdoors that only show up on specific inputs. See supply chain attacks.

What is an AI bill of materials?

An inventory of the models, datasets, adapters and libraries an AI system depends on, with versions and sources, so you can respond quickly when one is found to be compromised.

Change log

  1. 13 November 2026First published.
1300 654 898