Supply Chain

The AI supply chain spans base models, fine-tunes, datasets, and the libraries that load them. Any of these can be tampered with — a typosquatted model, an unsafe deserialization payload in model weights, or a compromised dependency — leading to code execution or subtle integrity failures.

How it shows up

• A typosquatted model ships a custom tokenizer that runs code on load via unsafe deserialization.
• A compromised Python package silently modifies inference behavior.
• A pre-trained model contains an undocumented backdoor trigger.

Mitigations

• Verify model and dataset provenance; prefer signed artifacts and trusted registries.
• Pin dependencies and generate an SBOM for your AI stack.
• Load models in sandboxes and avoid unsafe deserialization formats.
• Scan third-party models and datasets before adoption.

Related threats