Trained model artifacts — weights and code that carry their own supply-chain risks.
AI models are software artifacts too: weight files, tokenizers, and loading code pulled from hubs and registries. They carry novel risks — pickled weight files can execute arbitrary code on load, and a poisoned model can behave adversarially in ways no unit test catches.
Safeguard verifies model artifacts alongside traditional packages: scanning loaders for unsafe deserialization, checking origins, and tracking advisories that affect model-serving stacks.