An AI bill of materials extends the software bill of materials to the parts of an AI system that are not code. Alongside libraries and versions it records the base model and its provider, the datasets used to train or fine-tune it, any adapters loaded on top, the document sources a retrieval system searches, and the tools and connectors the model can call. CycloneDX and SPDX 3.0 both include formats for describing models and datasets.
The value is the same as an SBOM’s. When a model provider discloses a problem, a dataset is found to be poisoned or a connector turns malicious, the organisation can see within minutes which of its systems are affected. Without it, the model often appears in the vendor register as a single line and the other layers are invisible.
An AIBOM records provenance. It cannot prove a model is safe, because weights cannot be inspected for a hidden backdoor. Pair it with limits on what the model is allowed to do, so that a gap in the record does not become a gap in control.