An AI BOM (AI bill of materials, also written AIBOM, AI-SBOM or ML-BOM) is a machine-readable inventory of the models, datasets, training configuration, licences and AI dependencies inside a system, usually written in CycloneDX. To generate one, use the open-source OWASP AIBOM Generator for Hugging Face models, or Snyk’s snyk aibom or Cisco’s cisco-aibom to scan a codebase.
A software bill of materials answers the question “what is inside this application?” by enumerating every dependency and its provenance. An AI BOM answers the harder question for AI systems: what models am I shipping, where did their weights originate, which datasets trained them, and what licenses and risks do they carry? When a backdoored model can masquerade as a legitimate open-weight release on a public hub, that inventory stops being a compliance nicety and becomes a supply-chain control.
AI BOM Software Compared
AI and machine learning BOM software splits by where it looks: the model registry, the source code, the compiled artifact, or the cloud account. The details below were checked against each project’s or vendor’s documentation in October 2026.
| Tool | Where it looks | Gen AI and agent coverage | Output | Licence or access |
|---|---|---|---|---|
| OWASP AIBOM Generator | One Hugging Face model ID | Model metadata and model card | CycloneDX 1.6 JSON plus a completeness score | Apache-2.0, free |
Snyk AI-BOM (snyk aibom) | Local Python, Java, JavaScript and Go projects | Models, datasets, AI libraries and LLM API calls found in code | CycloneDX 1.6 JSON, optional HTML view | Snyk CLI v1.1298.3 or later |
Cisco AI BOM (cisco-aibom) | Code in seven languages, container images, cloud environments | Agents, tools, MCP servers and clients, prompts, guardrails, secrets | CycloneDX 1.6, SPDX 3.0, SARIF 2.1.0, JSON, HTML, CSV and more | Apache-2.0, free; needs an LLM |
| ReversingLabs Spectra Assure | Compiled packages and model files, no source needed | Identifies 8,000+ public models | CycloneDX 1.6, SPDX 3.0 | Commercial |
| HiddenLayer Model Scanner | Each model file it scans | Model Genealogy from the computational graph | CycloneDX | Commercial |
| Wiz AI-SPM | Cloud accounts, agentless | AI services, model usage, supporting infrastructure | Live inventory; no export format stated | Commercial |
A gen AI BOM has a different shape from a classic ML-BOM. A fine-tuned classifier is one weights file and one training set; an LLM application is a chain of hosted model calls, prompts, agent tools and MCP connections that never appear in a requirements file. Registry-side generators cannot see that chain, so start with the code-side scanners (Snyk, Cisco), then put runtime controls on the agents they surface, compared in our AI agent security tools guide.
The Standard: OWASP CycloneDX ML-BOM
Most AIBOM tooling converges on one format, so it is worth understanding it first. CycloneDX is an OWASP project and a full-stack bill-of-materials standard. It introduced Machine Learning transparency (ML-BOM) in version 1.5, released in June 2023, giving the format a structured way to represent models, datasets, and their configurations alongside conventional software components.
An ML-BOM extends a standard SBOM with the things that actually shape a model’s behavior and risk: model architecture and provenance, training and inference configuration, the datasets involved, and the documentation (model card) that describes intended use and limitations. In the CycloneDX 1.6 schema that means two dedicated component types, machine-learning-model and data, plus a modelCard object on each model. CycloneDX has since been ratified as an international standard, ECMA-424, with v1.6 as the 1st Edition (2024) and v1.7 as the 2nd Edition (December 2025). That matters for procurement and audit, because “give us a CycloneDX AIBOM” is a far cleaner contractual ask than “give us a spreadsheet of your models.” The open-source generators and ReversingLabs all write 1.6 today, so target 1.6 unless a consumer asks for 1.7.
SPDX, the other major SBOM standard, covers the same ground through the AI profile in SPDX 3.0. Its AIPackage class carries fields such as typeOfModel, informationAboutTraining, energyConsumption, safetyRiskAssessment and useSensitivePersonalInformation. Cisco AI BOM and ReversingLabs emit both formats.
Open-Source AIBOM Generators
These tools generate a bill of materials from models or repositories. They are the right starting point for teams that want transparency without a commercial contract.
OWASP AIBOM Generator is the standards-body option. Originally created by Helen Oakley and Dmitry Raidman and introduced at RSAC 2025, it was contributed to OWASP and now lives under the OWASP Gen AI Security Project. It generates an AIBOM for models hosted on Hugging Face, pulling metadata from the model card, configuration and repository files and producing downloadable CycloneDX 1.6 JSON, with visualization of model metadata and dependency information. The Apache-2.0 code runs as a web app, a CLI and an API, and a hosted version sits on Hugging Face Spaces. Every output gets a completeness score weighted across five categories: required CycloneDX fields (20%), AI metadata and provenance (20%), basic component identification (20%), the model card (30%) and external references (10%). Best for: teams that want a vendor-neutral, standards-aligned AIBOM for individual Hugging Face models.
Snyk AI-BOM comes at the problem from the codebase, not the model registry. It began as an experimental Snyk Labs tool and now ships as the snyk aibom command in the Snyk CLI, available from v1.1298.3, with the --experimental flag no longer required from v1.1304.0. It scans local Python, Java, JavaScript and Go projects for AI components — models, datasets, libraries, and even agent frameworks — detecting usage patterns such as Hugging Face transformers, OpenAI API calls, and training datasets, and emits CycloneDX v1.6 JSON. It also runs as an API. Its differentiator is that it leans on Snyk’s DeepCode engine to detect embedded AI usage in code even when there is no clear manifest or package reference — the opaque API call that traditional software composition analysis tools miss entirely. Best for: engineering organizations that want to discover AI usage across many repos rather than document one model at a time.
Cisco AI BOM is the most agent-aware of the three. Cisco released it in February 2026 both inside its commercial AI Defense product and as a standalone Apache-2.0 CLI. It scans codebases, container images and cloud environments for 30 AI component types, from models and datasets to agents, tools, MCP servers and clients, prompts, guardrails and secrets. It reads Python, JavaScript/TypeScript, Java, Go, Rust, Ruby and C#, with the deepest code-level detection in Python. The catch is that cisco-aibom analyze always requires an LLM for its classification stage, so decide where that step runs; a local Ollama model is a supported option. Best for: teams inventorying generative AI applications and agents rather than standalone models.
The approaches are complementary: a registry-side generator documents a model you have chosen, while a code-side scanner finds the models and providers your code actually touches, including the ones nobody told the security team about.
Commercial Platforms That Produce an AIBOM
For teams that want the bill of materials as one output of deeper security analysis, not the whole product, two model-security platforms build the inventory while inspecting artifacts for tampering, and one cloud platform keeps a live version.
ReversingLabs Spectra Assure added ML-BOM to its Spectra Assure software-supply-chain platform as part of its CycloneDX Extended BOM (xBOM) support, which also spans CBOM and SaaSBOM. Its ML-BOM provides visibility into every ML model in an environment and can identify over 8,000 publicly available models from sources like Hugging Face — even without source-code access. The platform performs deep model inspection across TensorFlow, PyTorch, scikit-learn, Keras, and NumPy, extracts and enriches model-card metadata, and exports in CycloneDX 1.6 and SPDX 3.0. The security weight here is in the analysis: unpacking ML model files out of complex software packages, binary-level inspection for hidden threats, and behavioral analysis to flag anomalous implants — the kind of inspection that surfaced BadSeek, a malicious model that masqueraded as the open-source Qwen2.5-Coder-7B-Instruct model after being backdoored. Best for: organizations that need supply-chain assurance on compiled commercial software and third-party models.
HiddenLayer generates an AIBOM automatically for every model its Model Scanner inspects, providing an auditable inventory of model components, datasets, and dependencies, exported in CycloneDX for supply-chain audits and licensing enforcement. It pairs the AIBOM with Model Genealogy, which analyzes a model’s computational graph to reveal architecture, origin, and intended function. We cover HiddenLayer’s broader posture role in our AI security posture management tools guide. Best for: teams that want the bill of materials tied to deep model-layer threat detection rather than as a standalone artifact.
Wiz takes the cloud-side view. Its AI-BOM is built on agentless cloud visibility and updates as teams deploy new services. It is a live inventory, not a portable file, and Wiz’s AI-BOM page names no export format, so confirm CycloneDX export before an audit. Best for: cloud-heavy estates that already run Wiz.
AI BOM Template (CycloneDX 1.6)
A usable AI BOM template is a CycloneDX 1.6 document with one machine-learning-model component per model and a modelCard holding the AI-specific fields. Field names follow the CycloneDX 1.6 JSON schema; replace the placeholders.
{
"bomFormat": "CycloneDX",
"specVersion": "1.6",
"serialNumber": "urn:uuid:<uuid>",
"version": 1,
"metadata": {
"component": { "type": "application", "name": "<your-ai-app>", "version": "<release>" }
},
"components": [
{
"type": "machine-learning-model",
"bom-ref": "model-1",
"name": "<owner/model>",
"version": "<revision or commit hash>",
"licenses": [{ "license": { "id": "<SPDX licence ID>" } }],
"externalReferences": [{ "type": "distribution", "url": "<download URL>" }],
"modelCard": {
"modelParameters": {
"approach": { "type": "supervised" },
"task": "<task>",
"architectureFamily": "<architecture family>",
"modelArchitecture": "<architecture name>",
"datasets": [
{ "type": "dataset", "name": "<training set>", "classification": "<public or internal>",
"contents": { "url": "<dataset URL>" } }
],
"inputs": [{ "format": "string" }],
"outputs": [{ "format": "string" }]
},
"quantitativeAnalysis": {
"performanceMetrics": [{ "type": "<metric>", "value": "<value>", "slice": "<eval split>" }]
},
"considerations": {
"users": ["<who uses it>"],
"useCases": ["<approved use>"],
"technicalLimitations": ["<known limitation>"],
"ethicalConsiderations": [{ "name": "<risk>", "mitigationStrategy": "<control>" }]
}
}
}
]
}
approach.type accepts supervised, unsupervised, reinforcement-learning, semi-supervised or self-supervised. Each datasets entry is either inline, as above, or a { "ref": ... } pointer to a separate component of type data.
The considerations block is the part you will write by hand. A July 2026 study of 97,940 AIBOMs produced by the OWASP generator from Hugging Face models found a mean completeness score of 54.31 out of 100. Licence data appeared in 73.14% of them, while fields such as intendedUse, ethicalConsiderations and modelExplainability were absent from every one. Generate the skeleton automatically, then have the model owner fill in the model card.
For a completeness check beyond the schema, use the seven clusters in the G7 Cybersecurity Working Group’s SBOM for AI – Minimum Elements, published on 12 May 2026 by agencies including CISA and Germany’s BSI (Infosecurity Magazine): metadata, system-level properties, models, dataset properties, key performance indicators, infrastructure and security properties. Its authors state the elements “are not mandatory”. Anything your template has no field for can go in properties as name-value pairs.
How to Generate an AI BOM
Pick the generator by what you are documenting. Commands are from each project’s documentation as of October 2026.
-
One Hugging Face model (OWASP AIBOM Generator). Paste the model ID (
owner/model) into the hosted generator at owasp-genai-aibom.org, or run it locally:git clone https://github.com/GenAI-Security-Project/aibom-generator cd aibom-generator pip install -r requirements.txt python3 -m src.cli google-bert/bert-base-uncased --output my_aibom.jsonWithout
--output, files land insboms/. For automation,POST /api/generate-with-reportreturns the BOM plus its completeness score and accepts an optional Hugging Facehf_token. -
A Python, Java, JavaScript or Go repository (Snyk). From the project root:
snyk aibom --json-file-output=aibom.json snyk aibom --html > aibom.htmlExit code 0 is success, 2 a failed run and 3 no supported files, so CI can tell “no AI here” from “the scan broke.”
--upload(with--repo) saves the result to your Snyk organization;snyk aibom testchecks it against tenant policies. -
An LLM app, agent codebase or container image (Cisco). Python 3.11+ and uv are prerequisites:
uv tool install --python 3.13 "cisco-aibom[agentic,llm-openai]" cisco-aibom analyze /path/to/project -o json -O report.json \ --llm-model <model-name> --llm-api-key $OPENAI_API_KEYSwap
-o jsonfor-o htmlto get a dashboard. Use thellm-awsextra instead for AWS Bedrock. -
Store and regenerate. Commit the BOM beside the release it describes and regenerate it whenever a model, dataset or provider changes.
Why the AIBOM Matters: Model Serialization Attacks
The reason an AIBOM is a security control and not just paperwork comes down to how models are distributed. Many model formats are derived from Python’s pickle module, and a model serialization attack embeds malicious code into a model file during saving, so it executes the moment the file is loaded — a modern Trojan horse. This is the threat that maps to the supply-chain risks in the OWASP Top 10 for LLM Applications.
The scanners that answer this are compared in full — by format, by evasion resistance, and by where the gate belongs in the pipeline — in the AI model scanning tools guide. The standard open-source defense is ModelScan from Protect AI (Apache-2.0 licensed; Protect AI was acquired by Palo Alto Networks in 2025). ModelScan reads a model file’s contents looking for unsafe code signatures, supports pickle-derived formats (PyTorch, scikit-learn, XGBoost), TensorFlow SavedModel, and HDF5/Keras, and ranks findings as critical, high, medium, or low. It outputs to console or JSON, and its exit codes make it straightforward to gate a CI pipeline on a clean scan. An AIBOM tells you which models you have; a scanner like ModelScan tells you whether any of them are unsafe to load. Run both. (The safer long-term fix is the safetensors format, designed by Hugging Face so that it is structurally impossible to embed executable code in a model file.)
How to Build an AIBOM Practice
For a team starting from zero:
- Pick the format first. Standardize on CycloneDX ML-BOM. It is an OWASP standard, ECMA-ratified, and supported by every generator above, which keeps your inventory portable.
- Generate at two layers. Use a code-side scanner (Snyk AI-BOM or Cisco AI BOM) to discover AI usage across repositories, and a registry-side generator (OWASP AIBOM Generator) to document each external model you adopt.
- Pair inventory with inspection. Gate model ingestion on a ModelScan pass in CI so a backdoored or serialization-attacked model never reaches your registry.
- Escalate to a platform when scale demands it. When you are ingesting models continuously or need assurance on third-party compiled software, a commercial platform (ReversingLabs, HiddenLayer) adds binary-level inspection and behavioral analysis the lightweight generators do not attempt.
The AIBOM is becoming a procurement requirement the same way the SBOM did before it — auditors and downstream customers will increasingly ask “what is in your AI?” The teams that can answer with a signed, standards-formatted bill of materials, backed by a scan that proves the artifacts are clean, will move through those conversations far faster than the ones reconstructing the answer after the fact. For the broader visibility discipline these tools feed into, see our AI security posture management tools guide; for catching tampering before deployment, the LLM vulnerability scanner comparison at bestllmscanners.com. Proving where a model came from after it leaves your registry is the job of AI model watermarking tools, another layer of a broader AI security tools program.
FAQ
What is an AI-SBOM (also called an AIBOM or ML-BOM)?
An AI bill of materials is a machine-readable inventory of an AI system’s components — the models, the datasets, the training and inference configuration, and the model-card documentation — recorded with their provenance and licences. AI-SBOM, AIBOM, and ML-BOM name the same idea; CycloneDX calls it an ML-BOM.
How is an AI-SBOM different from a regular SBOM?
A conventional SBOM enumerates software dependencies and their versions. An ML-BOM extends that with the artifacts that actually shape a model’s behaviour and risk — architecture and provenance, the datasets involved, training and inference configuration, and the model card — none of which a code-only SBOM has a place to record.
What format should an AI-SBOM use?
CycloneDX ML-BOM is the de-facto standard: it is an OWASP project, ratified internationally as ECMA-424, and emitted by every open-source generator compared above, which keeps the inventory portable and clean to hand to an auditor. SPDX 3.0 is the alternative, with a dedicated AI profile, and Cisco and ReversingLabs can emit both.
Is there a free AI BOM generator?
Yes. The OWASP AIBOM Generator and Cisco’s cisco-aibom are both Apache-2.0 licensed. OWASP’s runs in a browser against a Hugging Face model ID and scores the result for completeness. Cisco’s scans source code and container images for models, agents, MCP servers and prompts, though its analysis step needs an LLM, hosted or local.
Is an AI-SBOM required by regulation?
No universal legal mandate exists today. The G7’s SBOM for AI – Minimum Elements, published on 12 May 2026 by agencies including CISA and the UK NCSC, states that its elements are not mandatory. Customers and auditors ask regardless, so map the BOM to the AI security audit frameworks you already report against.