AI Supply Chain Risk
AI supply chain risk is the danger that the models, datasets, libraries, plugins and AI vendors a business relies on are vulnerable, tampered with or mishandle data, bringing problems into its own AI systems.
Also called: AI supply chain security, model supply chain, AI bill of materials, AIBOM, malicious models
What is AI supply chain risk?
Few businesses build AI from scratch. They use foundation models from AI companies, download open models from public hubs, fine-tune them with public datasets, rely on open-source AI libraries, add plugins and connectors, and buy AI features from software vendors. Each of these is a link in the AI supply chain, and a weakness in any of them can become yours. It's the AI version of a traditional software supply chain attack.
OWASP ranks supply chain fourth in the 2026 edition of its Top 10 for LLM applications.
The main risks
OWASP's guidance on LLM supply chain risks includes:
- Vulnerable components: outdated or flawed AI libraries and frameworks, just like other open-source dependencies.
- Vulnerable or malicious pre-trained models. OWASP notes that models are "binary black boxes": unlike code, inspecting them offers little assurance. A model can contain hidden backdoors or biases. Some model file formats can also run code when loaded, so a downloaded model can carry malware.
- Poisoned datasets used for training or fine-tuning. See AI data poisoning.
- Outdated or deprecated models that no longer get security fixes.
- Licensing risks, where model or dataset licenses restrict how you can use them.
- Risky plugins, extensions and connectors that gain access to your data.
- AI vendors' data practices: where your prompts and files go, how long they're kept and whether they're used for training.
How to manage it
- Inventory your AI components: models, versions, datasets, libraries and AI-enabled products. An AI bill of materials (AIBOM) extends the SBOM idea to models and data.
- Use trusted sources. Prefer established providers and verified publishers, and check integrity when downloading models.
- Use safe model formats that don't execute code on loading, and scan model files.
- Pin and patch AI libraries like any other dependency.
- Review AI vendors as part of third-party risk management: data handling, training opt-outs, security certifications and breach notification.
- Test models before deploying them and after updates.
- Vet plugins and connectors, and restrict who can install them.
Related terms
Sources
- LLM03:2025 Supply Chain (OWASP Gen AI Security Project)
- AI Risk Management Framework (NIST)
Drafted with AI assistance, then edited and fact-checked by a person. How we use AI.
