An AI bill of materials (AIBOM) is an inventory of the models, datasets, dependencies and configuration that make up an AI system, in a machine-readable format. It extends the software bill of materials idea to machine learning so you can trace provenance, assess risk and answer customer or regulator questions.
Key takeaways
- An AIBOM records models and datasets, not just code packages.
- Two formats cover it: CycloneDX ML-BOM and the SPDX 3.0 AI profile.
- Start by listing every model and dataset in production, with source and license.
- It supports risk checks for bias, data integrity and model security.
What goes in an AIBOM?
CycloneDX describes its ML-BOM as covering models and datasets with their dependencies, dataset provenance and ethical considerations, training methodologies and the configuration of AI frameworks. It says this supports assessing risks related to bias, data integrity and model security.
A practical record per model includes name, version, source, license, training data references, and where it is deployed.
Which formats exist?
CycloneDX offers ML-BOM. SPDX 3.0 has an AI profile, whose properties include autonomyType, domain, energyConsumption, hyperparameter, limitation, safetyRiskAssessment, standardCompliance, typeOfModel and useSensitivePersonalInformation, according to the specification navigation. The profile requires a licensing relationship for every AIPackage.
How do you create one?
- List models: hosted APIs, open weights and fine-tunes.
- List datasets, including retrieval data.
- Record versions and sources, such as a model hub URL or vendor.
- Capture licenses and known limitations.
- Export to CycloneDX or SPDX and store it with each release.
Try the AI BOM demo, and generate a normal SBOM with the SBOM demo. Safeguard also offers a free browser dependency scan at gold.safeguard.sh/scan for the code packages around your models.
What does a minimal entry look like?
Keep the first version small. One row per model is enough to answer most questions:
| Field | Example value |
|---|---|
| Name and version | The model name and the exact version or revision you load |
| Source | Vendor API or model hub page |
| License | The license text that applies to the weights or API |
| Training data | Reference to datasets, if known and disclosed |
| Known limitations | What the provider says it should not be used for |
| Where used | The services and features that call it |
Gaps are normal at first. Mark unknown fields as unknown instead of guessing, then close them over time.
Why does it matter?
An inventory lets you answer which systems use a compromised or recalled model. It also supports governance work, see how to start an AI governance framework, and EU AI Act record keeping.
Frequently asked questions
What is the difference between SBOM and AIBOM?
An SBOM lists software components. An AIBOM adds models, datasets and training details.
Is an AIBOM required by law?
We found no source that requires one by that name. It can support documentation duties, so check with counsel.
Is ML-BOM the same as AIBOM?
ML-BOM is CycloneDX's name for the capability, and AIBOM is the general term.
Who maintains the AIBOM?
The team that ships the model, updated at each release.
Sources
Last reviewed: 19 October 2026.
