Skip to content
All posts

AIBOM Parameters: A Field-by-Field Checklist

October 2, 2026 · PlayCISO
TL;DR · the short version

An AIBOM is only useful if it captures the right fields, in six groups: models (each base model and fine-tune with version, source, license, model card, architecture, parameter count); datasets (source, license, provenance, sensitive-data flags); AI/ML code dependencies pinned to versions and licenses; configuration and runtime; risk and assurance metadata (eval metrics, bias, known CVEs, signing); and document metadata in a machine-readable standard (CycloneDX ML-BOM or SPDX 3.0). This post gives a copy-paste checklist, free on PlayCISO.

Put it into practice
MCP Guard
Audit an MCP / agent configuration for plaintext secrets, unpinned installs, shell pipes and over-broad roots.
Open the tool →
AIBOM — AI Bill of Materials: models, datasets and dependencies inventoried with license and provenance

An AIBOM (AI Bill of Materials) is only as useful as the fields it captures. A document that lists "we use an LLM" is theatre; one that names every model, dataset, dependency, license and known risk is an asset you can audit, hand to a customer, or file with a regulator. This is the field-by-field checklist of what to put in one — and why each parameter earns its place — mapped to the two standards that matter, CycloneDX ML-BOM and the SPDX 3.0 AI Profile.

New here? Start with what an AIBOM is and why it matters, and the SBOM vs AIBOM vs MLBOM distinction. This post is the "what fields" companion to our how to build an AIBOM walkthrough.

Why the parameters matter

AIBOMs crossed from research artifact to procurement requirement: enterprise buyers now ask for one in due diligence, and high-risk-system documentation under regimes like the EU AI Act expects equivalent content. The value is entirely in the fields. A missing license field is how a non-commercial or copyleft model licence silently ends up in a paid product. A missing dataset provenance field is how training data you cannot legally use gets baked into a model. A missing version field is how you fail to answer "are we affected?" when a model or library has a disclosed vulnerability. Each parameter below exists because its absence has already burned someone.

The parameters, field by field

1. Models — every one, as a first-class component

Each base model and each fine-tune is its own component. Capture:

  • Name & version — the exact model and revision/commit, so "are we affected?" has an answer.
  • Source / registry — where it came from (Hugging Face repo, internal registry, vendor API), the single most important provenance signal.
  • License — the model licence and any acceptable-use terms. This is the field that catches the silent licence trap.
  • Model card / intended use — purpose, intended and out-of-scope uses.
  • Architecture & size — family, parameter count, modality.
  • Relationship — for a fine-tune, which base model it derives from. AIBOM identity, provenance and lineage covers how to record it.

2. Datasets — training and evaluation

  • Name & source — each training and evaluation dataset and where it came from.
  • License & usage rights — whether you may use it the way you are using it.
  • Provenance — origin and collection method; synthetic vs. scraped vs. licensed.
  • Sensitivity flags — PII, regulated or copyrighted content present in the data. The AIBOM data section fills in each of these for a worked example.

3. AI/ML code dependencies — the SBOM layer

An AIBOM still contains a normal SBOM underneath it: the frameworks and libraries (the ML stack, serving layer, orchestration) pinned to versions and licenses, so known-vulnerability and licence checks work exactly as they do for any software. Our AI Dependency Scanner → reads these from a manifest.

Useful next, if this was relevant to you

Showing 1 of 3: AI Security Risk Mapper

4. Configuration & runtime

  • Training configuration — method (pre-train, fine-tune, LoRA, RAG), key hyperparameters where disclosable.
  • Inference configuration — serving setup, context limits, tool/function access the model is granted. The AIBOM security and runtime section covers tools, MCP servers and RAG indexes.
  • Deployment context — where it actually runs (cloud, on-prem, edge), because a model with tool access in production is a different risk than the same weights on a laptop.

5. Risk & assurance metadata

  • Performance / evaluation metrics — how it was measured, on what. See AIBOM evaluation, intended use, limitations and risk.
  • Bias & safety considerations — documented limitations and known failure modes.
  • Known vulnerabilities (CVEs) — against the model, its loader, and its dependencies.
  • Provenance & integrity — signing / attestation and a hash of the weights, so you can prove what you shipped. Pair this with a scan: our Model Risk Scanner → reads a Hugging Face model against risk policies before it enters your environment.

6. Document metadata & format

  • Author, timestamp, version of the AIBOM itself, and the supplier it describes.
  • A machine-readable standard — CycloneDX ML-BOM (machine-learning-model + data component types; common in security/CI-CD tooling) or the SPDX 3.0 AI Profile (model type, training methods, data handling, explainability, limitations, energy use; Linux-Foundation/ISO-rooted, often preferred for regulatory filings). A PDF is not an AIBOM — your tools cannot read it.

The copy-paste AIBOM checklist

Run this against any AIBOM you produce or receive:

  • ☐ Every base model and fine-tune listed as its own component (name + version + source)
  • ☐ A license on every model and every dataset
  • ☐ Provenance for each model (registry/repo) and dataset (origin + collection method)
  • ☐ Sensitive-data flags on datasets (PII / regulated / copyrighted)
  • ☐ AI/ML code dependencies pinned to versions and licenses
  • ☐ Training and inference configuration, and where the model runs
  • ☐ Evaluation metrics and documented bias/safety limitations
  • ☐ Known CVEs for the model, loader and dependencies
  • ☐ Integrity: a weights hash and signature/attestation
  • ☐ Document metadata (author, timestamp, supplier) in CycloneDX ML-BOM or SPDX 3.0 AI Profile — not a PDF

Generate and check an AIBOM on PlayCISO

You do not need a platform licence to start. These AIBOM tools are part of PlayCISO Pro, and a free account includes 3 Pro tool runs:

Open-source examples to learn from (GitHub)

If you want real, working references for the fields above:

  • OWASP AIBOM Generator — extracts model metadata from Hugging Face into a CycloneDX AIBOM and scores its completeness (a good model for which fields to fill).
  • cdxgen (CycloneDX) — generates BOMs across many ecosystems and integrates into CI/CD and Dependency-Track.
  • CycloneDX Tool Center and SPDX — the specifications themselves, which define every field precisely.

An AIBOM without these fields is a document; an AIBOM with them is a control. Capture the parameters, emit them in a machine-readable standard, and you can answer every supply-chain question a customer, auditor or incident will ever ask about your AI. Next: compare the AIBOM tools and vendors that produce them, and read what an ML-BOM is for the model-specific layer.

Build your first one now with the AI BOM generator, validate it with the AIBOM Validator, and keep up with AI supply-chain security in PlayCISO AI Labs.

Frequently asked questions

What fields must an AIBOM contain?

At minimum: every model (base and fine-tune) as a first-class component with version, source/registry, license and model card; every training and evaluation dataset with source, license and provenance; the AI/ML code dependencies with versions and licenses; the training and inference configuration; risk metadata (performance/eval metrics, bias and safety notes, known CVEs, provenance/signing); and document-level metadata in a machine-readable format (CycloneDX ML-BOM or SPDX 3.0 AI Profile).

Which format should an AIBOM use — CycloneDX or SPDX?

CycloneDX ML-BOM uses machine-learning-model and data component types and is common in security and CI/CD tooling. The SPDX 3.0 AI Profile (plus Dataset Profile) describes model type, training methods, data handling, explainability, limitations and even energy use, and as a Linux Foundation / ISO-rooted standard is often preferred for regulatory filings. Both are stable and machine-readable; pick for your audience, and convert between them when needed.

What is the difference between an AIBOM and an SBOM?

An SBOM inventories code. An AIBOM extends it with the components an SBOM never described — the models and the datasets, each with its own license and provenance — because in an AI product the model and its training data are risk-bearing dependencies just like an npm package. See our SBOM vs AIBOM vs MLBOM explainer for the full comparison.

How do I generate and check an AIBOM?

On PlayCISO you can generate a CycloneDX-style AIBOM from a manifest with the AI BOM generator, convert an ML-BOM with the ML-BOM Converter, validate structure with the AIBOM Validator, start from the AIBOM Template, diff two AIBOMs with AIBOM Diff, and check model and dependency licenses with the Model License Checker and AI Dependency Scanner. The AI BOM generator, converter, validator, template, diff and licence checker are all part of Pro (a free account gets 3 Pro tool runs).

Ready to practise the decisions these articles describe?

Run your first Cyber War Room free →

Pro