Best AI Security Tools
Isometric network of red hexagonal nodes linked by thin lines to translucent purple cubes on glowing platforms, evoking ML model component dependencies
Tools

ML BOM Software: Tools to Generate, Sign and Store ML-BOMs

ML BOM software: which free and commercial tools generate, validate, sign and store CycloneDX ML-BOMs, and the spec-version gaps to plan around.

By Best AI Security Tools Editorial · · 8 min read

ML BOM software covers four jobs: generating a machine learning bill of materials, validating it against the schema, signing it, and storing it where a policy engine can read it. For most teams the pick is an open-source CycloneDX stack: cdxgen to generate, CycloneDX CLI to validate, cosign to sign, Dependency-Track to store. Commercial platforms add binary-level model inspection on top.

An ML-BOM extends a software bill of materials with what shapes a model’s behaviour: the model, its datasets, training configuration and provenance. CycloneDX added ML-BOM in version 1.5, released 26 June 2023, and OWASP’s LLM03:2025 Supply Chain entry says AI BOMs and ML SBOMs “are an emerging area and you should evaluate options starting with OWASP CycloneDX.” The threat behind that advice is MITRE ATLAS AML.T0010 AI Supply Chain Compromise, sub-technique AML.T0010.003 Model, per the ATLAS 2026.09 data. If you came here for multi-level bills of materials in manufacturing ERP, that is a different product category.

What does ML BOM software actually do?

It turns models, datasets and AI dependencies into a machine-readable inventory, then makes that inventory trustworthy and queryable. That splits into generation, schema validation, signing, and storage with policy checks. Commercial platforms bundle some of these jobs, but no tool below covers all four, so most teams assemble a stack. Details were checked against each project’s documentation in October 2026.

ToolJobInputOutputLicence or access
cdxgen (aibom, -t ai)GenerateSource trees, Hugging Face purls and URLs, GGUF files, ModelfilesCycloneDX 1.7 by default, 1.6 on requestApache-2.0
OWASP AIBOM GeneratorGenerate, scoreOne Hugging Face model IDCycloneDX 1.6 JSON plus completeness scoreApache-2.0
AIBoMGen CLIGenerate, validate, scoreRepos scanned for Hugging Face usage, model IDsCycloneDX 1.7 by default, JSON or XMLAGPL-3.0
Snyk (snyk aibom)Generate, policy testPython, Java, JavaScript and Go projectsCycloneDX 1.6 JSONSnyk CLI v1.1298.3 or later
Cisco AI BOMGenerateCode in seven languages, container images, cloud environmentsCycloneDX 1.6, SPDX 3.0, SARIF and othersApache-2.0; needs an LLM
ReversingLabs Spectra AssureGenerate, inspectCompiled packages and model filesCycloneDX 1.6, SPDX 3.0Commercial
CycloneDX CLIValidate, merge, diff, signAny CycloneDX BOMSchema verdict, v1.0 to v1.7Open source
cosignAttestBOM plus container imageSigned attestation via --type cyclonedxOpen source
Dependency-TrackStore, trackCycloneDX BOMsComponent inventory and risk trackingOpen source
mlbomdocReadCycloneDX ML-BOM JSONConsole, JSON, Markdown, PDFApache-2.0

Which ML BOM generator fits your inputs?

Pick by what you hold. A Hugging Face model ID suits the OWASP AIBOM Generator or AIBoMGen CLI. A source repository suits cdxgen, Snyk or Cisco. A compiled vendor package with no source suits ReversingLabs. Teams shipping LLM applications need a code-side scanner, because hosted model calls and agent tools never appear in a dependency manifest.

cdxgen is the broadest free generator. It is an OWASP Foundation production project, and its AI-BOM mode accepts a directory, a pkg:huggingface/ purl or Hub URL, a local GGUF file or a Modelfile. It writes models as machine-learning-model components and datasets as data components, fills modelCard with task, datasets and input/output shape, and picks up AI services, agent frameworks, prompt configs and MCP surfaces in source trees. A companion cdx-audit command re-audits a saved BOM against AI governance, AI security, agent and MCP-server categories.

AIBoMGen CLI is a Go tool from IDLab (Ghent University and imec), listed in the CycloneDX Tool Center. It walks a repository’s code and config files for Hugging Face usage and emits one AIBOM per model. Its validate command checks the official CycloneDX schemas for 1.2 to 1.7 and can enforce a minimum completeness score, while vuln-scan pulls the Hugging Face Hub’s per-file scan results (ClamAV, Protect AI, VirusTotal and others) into the BOM. The AGPL-3.0 licence needs legal review before you redistribute it.

The OWASP AIBOM Generator documents one Hugging Face model at a time and runs as a web app, CLI or API. Snyk’s snyk aibom maps models, agents, tools and MCP components, and snyk aibom test checks the result against tenant policy. Cisco AI BOM covers 30 component types including prompts, guardrails and secrets, but cisco-aibom analyze always requires an LLM. ReversingLabs Spectra Assure can identify over 8,000 public models without source access. Per-tool detail is in our AI-SBOM and ML-BOM tools comparison.

Where does ML BOM software sit in the pipeline?

At build time, between model intake and deployment, off the request path, so it adds no inference latency. Without it, a model pulled from a hub moves from download to fine-tuning to registry to serving image with nothing recording what it is. With it, each release carries a signed, schema-valid record that a deploy gate can check.

The controlled path runs in five steps:

  1. Intake. Pull the model at a pinned revision and scan the file before anything loads it. An ML-BOM records which model you have, not whether it executes code on load; see our AI model scanning tools guide for that layer.
  2. Generate the BOM from the repository that consumes the model.
  3. Validate and fail the build on schema errors.
  4. Attest the BOM to the serving image.
  5. Store and gate. Upload to the BOM store; the deploy step refuses images without a valid attestation.

Steps 2 to 4 in CI, with flags from each project’s documentation:

cdxgen -t ai --include-formulation --spec-version 1.6 -o ml-bom.json .
cyclonedx validate --input-file ml-bom.json --input-version v1_6 --fail-on-errors
cosign attest --predicate ml-bom.json --type cyclonedx --key cosign.key <image>

CycloneDX CLI also merges and diffs BOMs, which helps when a code-side and a registry-side generator each describe part of one release. cosign’s --type flag accepts cyclonedx alongside SPDX, SLSA provenance and OpenVEX predicates.

What are the trade-offs?

Three gaps decide how much value the stack delivers: generated BOMs are structurally complete but thin on AI-specific fields, tools disagree on the CycloneDX version, and the main open-source BOM store does not yet treat model components as first-class. None is a reason to skip ML-BOMs, but each needs a decision before rollout.

Completeness. A July 2026 measurement study of roughly 97.5K AIBOMs generated from public Hugging Face repositories found required fields fully represented, while model-card, responsible-use, environmental and limitation fields were weakly represented or missing. Automation produces the skeleton; the model owner still writes the considerations. Gate on a completeness score, such as AIBoMGen’s --min-score, not on schema validity alone.

Version skew. cdxgen and AIBoMGen default to CycloneDX 1.7; Snyk, Cisco, the OWASP generator and ReversingLabs emit 1.6. CycloneDX CLI validates against v1.7 unless told otherwise, so set --input-version to match. Dependency-Track added 1.6 import on the 4.11.4 milestone, and 1.7 import closed under the 5.1 milestone in August 2026. If your store is still on v4, pin generators to 1.6.

Consumer support. Dependency-Track’s request to add machine-learning-model and data classifiers is still open; one commenter reports that v4.13.0 saves the new classifiers but does not show them in the project overview. Treat it as the inventory for the software around the model, not as an ML-BOM viewer. For human review, mlbomdoc renders an ML-BOM to Markdown or PDF, though its latest release, 0.1.1, dates from July 2024.

Data handling and cost. Cisco’s analysis depends on an LLM, so route that step to a model your data-handling policy allows. The open-source tools cost nothing to license; the real cost is engineering time for the gates and for model owners completing model cards. The documentation cited here publishes no scan-time figures, so measure CI duration before making the gate blocking.

Who should pick which stack?

The open-source stack (cdxgen, CycloneDX CLI, cosign, Dependency-Track) fits these profiles:

  • Platform teams fine-tuning open-weight models from Hugging Face who need a per-release record of base model, revision and datasets.
  • Organisations already running Dependency-Track for software SBOMs that want ML components in the same inventory and can live with the classifier gap.

A code-side, agent-aware scanner (Snyk or Cisco AI BOM) fits these:

  • Teams whose AI estate is mostly hosted model APIs, agent frameworks and MCP servers rather than weights files. Pair the inventory with the runtime controls in our AI agent security tools guide.
  • Existing Snyk customers who want snyk aibom test policy checks beside their current scans.

ReversingLabs Spectra Assure fits third-party risk teams that receive compiled software and need to know which models ship inside it.

Who shouldn’t rely on ML BOM software alone:

  • Teams whose main exposure is malicious model files. The BOM names the model; a model-file scanner catches the payload.
  • Teams expecting the BOM to satisfy an auditor by itself. Map its fields to the AI security audit frameworks you already report against.
  • Teams with no source access and no budget for binary analysis, since registry-side generators only document models someone already knows about.

FAQ

what is ml bom software

ML BOM software generates, validates or stores a machine learning bill of materials: a machine-readable list of the models, datasets, training configuration and licences inside an AI system. Most tools write the CycloneDX ML-BOM format introduced in version 1.5, and some also export the SPDX 3.0 AI profile. Commercial platforms bundle the inventory with model inspection.

is there free ml bom software

Yes. cdxgen, the OWASP AIBOM Generator and Cisco AI BOM are Apache-2.0 licensed, and AIBoMGen CLI is AGPL-3.0. CycloneDX CLI, cosign and Dependency-Track cover validation, signing and storage at no licence cost. Cisco’s tool needs an LLM for analysis, which can add API spend depending on the model you configure.

what is the difference between an ml-bom and an sbom

An ML-BOM is an SBOM extended for machine learning. A conventional SBOM lists software packages and versions; an ML-BOM adds machine-learning-model and data components and a model card holding architecture, task, datasets and considerations. In CycloneDX both live in the same document, so one file can describe the application and its models.

can dependency-track import an ml-bom

Yes, as a CycloneDX file. The 4.11.4 milestone added CycloneDX 1.6 import, and 1.7 import closed under the 5.1 milestone. Model-specific handling is thinner: the request to add machine-learning-model and data classifiers remains open, so do not count on it to present model cards to reviewers. Render those with mlbomdoc instead.

Sources

  1. Machine Learning Bill of Materials (AI/ML-BOM) — CycloneDX
  2. Introducing OWASP CycloneDX v1.5 — CycloneDX
  3. LLM03:2025 Supply Chain — OWASP Gen AI Security Project
  4. ATLAS tactics, techniques and case studies data — MITRE ATLAS (GitHub)
  5. cdxgen — OWASP cdxgen (GitHub)
  6. AI-BOM documentation — cdxgen (GitHub)
  7. AIBoMGen CLI — IDLab (GitHub)
  8. CycloneDX Tool Center — CycloneDX
  9. OWASP AIBOM Generator — GenAI Security Project (GitHub)
  10. AIBOM command — Snyk CLI documentation
  11. Cisco AI BOM — Cisco AI Defense (GitHub)
  12. Secure Your AI Supply Chain with the ML-BOM — ReversingLabs
  13. CycloneDX CLI — CycloneDX (GitHub)
  14. cosign attest command reference — Sigstore (GitHub)
  15. Support Import of CycloneDX 1.6 BOMs, issue #3584 — Dependency-Track (GitHub)
  16. Support Import of CycloneDX 1.7 BOMs, issue #5818 — Dependency-Track (GitHub)
  17. Basic Support for Classifiers Added in CycloneDX 1.5 and 1.6, issue #4361 — Dependency-Track (GitHub)
  18. mlbomdoc — PyPI
  19. A Large-Scale Measurement of AI Bill of Materials Completeness in Hugging Face Models — arXiv
  20. AIPackage class — SPDX Specification 3.0.1
Subscribe

Best AI Security Tools — in your inbox

Comparing the AI security tooling landscape, with numbers — delivered when there's something worth your inbox.

No spam. Unsubscribe anytime.

Related