Skip to content
By 14 min readAI Security

The AI Bill of Materials: How to Inventory Your AI Supply Chain

An AI bill of materials inventories the models, datasets, model files, libraries and tools an AI system is built from. What CycloneDX, SPDX and the G7 ask for, and a one-week plan.

The AI Bill of Materials: How to Inventory Your AI Supply Chain, by Deepak Gupta on guptadeepak.com

An AI bill of materials (AI-BOM) is an inventory of what an AI system is built from. That means the base model and every fine-tune on top of it, the datasets, and the model files. It also covers the code that loads and serves them, and the outside services and tools the system calls. Two published formats can carry it today. CycloneDX has supported machine learning models and datasets since version 1.5, and SPDX 3.0 has dedicated AI and Dataset profiles.

No law I could find requires a document called an AI-BOM. The EU AI Act asks for the content, not the format. Since May 2026 there is also a G7 list of minimum elements that tells you what a reasonable one contains.

Verified as of 30 October 2026 against the CycloneDX 1.7 JSON schema, the SPDX 3.0.1 specification and the G7 minimum elements document. Also checked: the NSA, CISA and FBI information sheet on AI data security, and the Official Journal text of Regulation (EU) 2024/1689. Every claim below links to its source.

What is in an AI system's supply chain

A conventional SBOM (software bill of materials) lists libraries and their versions. An AI system has those too, plus several components a package manager never sees. For one product, the list usually looks like this:

  • Base models. The foundation model you call through an API or run from downloaded weights.
  • Fine-tunes, adapters and quantized copies. Each is a new artifact derived from a parent, and the parent's problems travel with it.
  • Datasets. Training, fine-tuning, evaluation, and the document collections you retrieve from at query time.
  • Model files and their formats. A weights file is not inert data in every format. Some formats execute code when loaded.
  • Inference dependencies. The framework that loads the model, the serving layer, and gateway libraries that route calls between providers.
  • Hosted model APIs. A model you never hold, identified only by a provider and a model ID.
  • Agent tools and MCP servers. Every tool an agent can call is code and data you did not write, reached over a connection you have to trust. MCP (Model Context Protocol) is the common standard for connecting those tools.

The question an AI-BOM answers is the one every incident response starts with. When a component turns out to be bad, which of our products contain it, and which version?

The documented failure modes

Each layer of that list has already failed in public, with a primary write-up to show for it.

Model files that run code when you load them

Hugging Face's own pickle scanning documentation states the problem directly. Pickle "is the default format for PyTorch model weights", the page says. It also warns: "There are dangerous arbitrary code execution attacks that can be perpetrated when you load a pickle file." The Hub scans uploads, and the same page says the scan "is not 100% foolproof."

Attackers have used this. In February 2024, JFrog reported around 100 malicious models on Hugging Face. One opened a reverse shell to a remote IP the moment it was loaded with torch.load(). In February 2025, ReversingLabs described two models that slipped past the Hub's scanner by using 7z compression and a deliberately broken pickle stream. The payload ran before the break. Hugging Face removed them within 24 hours and updated its scanner.

The safe-loading flag has failed too. PyTorch advisory CVE-2025-32434, rated Critical, showed that torch.load with weights_only=True could still lead to remote code execution in versions up to 2.5.1. The fix shipped in 2.6.0.

The alternative is a format that stores only tensors. Safetensors describes itself as "a new simple format for storing tensors safely (as opposed to pickle)." This is why the file format belongs in your inventory as a field of its own. A .safetensors file and a pickled .bin file can hold the same weights and carry very different risk.

Dependencies that get swapped or hijacked

Between 25 and 30 December 2022, PyTorch-nightly installs on Linux through pip pulled in a malicious torchtriton package. Someone had registered the same name on PyPI, and pip preferred it over PyTorch's own index. The binary read SSH keys, /etc/passwd and the first 1,000 files in the home directory, then sent them out over DNS. Stable releases were not affected.

In March 2026 it was a gateway library. LiteLLM versions 1.82.7 and 1.82.8 were published to PyPI with credential-stealing code. The maintainers' post-incident write-up says the packages were "live for about 40 minutes before being quarantined by PyPI." The root cause was an unpinned security scanner in their CI pipeline that exposed release credentials. Their incident thread notes that users of the proxy Docker image were not affected, because its dependencies were pinned. I covered one downstream breach and the wider pattern separately.

Forty minutes is the point. Nobody reviews a dependency in that window. What saves you is knowing, within the hour, which of your systems pulled that version.

Poisoned data you did not collect

The May 2025 joint information sheet AI Data Security, from NSA, CISA, FBI and partner agencies in Australia, New Zealand and the UK, names three risk areas: "data supply chain, maliciously modified ('poisoned') data, and data drift."

Its most concrete example is split-view poisoning. Curated web-scale datasets often point to content on domains that have since expired. Whoever buys the domain controls what the dataset now contains. The sheet puts the cost of effective poisoning at "roughly $1,000 USD" in many cases, and "as little as $60 USD" for one named dataset. Its mitigation is a supply chain control: attach a cryptographic hash to raw data, and check it at download.

For anyone building on a model they did not train, the sheet goes further. Foundation model developers "need to be able to provide assurances regarding the data and sources used," and consumers should "track the training data used in various model lineages." That is an AI-BOM requirement in everything but name.

Tools and MCP servers

An agent's tools are supply chain too. JFrog's write-up of CVE-2025-6514 (CVSS 9.6) showed that mcp-remote, a proxy that connects AI clients to remote MCP servers, could be made to execute operating system commands by a malicious server. It affected versions 0.0.5 to 0.1.15 and was fixed in 0.1.16. The trigger was simply connecting to the wrong server. An inventory that lists your MCP servers, their versions and their endpoints is how you find out whether you were exposed.

The standards that exist to inventory it

There are two machine-readable formats, one documentation convention, and one government list of what to put in them.

CycloneDX ML-BOM

The current release is CycloneDX 1.7, released 21 October 2025, which the project also lists as Ecma standard ECMA-424. Machine learning support, which CycloneDX calls ML-BOM, first appears in the 1.5 schema.

In the 1.7 schema, a model is a component of type machine-learning-model and a dataset is a component of type data. A model can carry a modelCard object with three main parts:

  • modelParameters: learning approach, task, architectureFamily, modelArchitecture, datasets, inputs and outputs.
  • quantitativeAnalysis: performanceMetrics.
  • considerations: users, useCases, technicalLimitations, ethicalConsiderations, fairnessAssessments and more.

Two general CycloneDX features do a lot of the AI work. Every component has hashes, which is where a weights file's digest goes. The pedigree object has ancestors, defined as the components a component "is derived from." I use that for base model to fine-tune lineage. Hosted model APIs and MCP servers fit the service object, which has endpoints, authenticated and trustZone fields.

SPDX 3.0 AI and Dataset profiles

SPDX's published specification is version 3.0.1. Its AI Profile is "designed to provide a standardized way of documenting and sharing information about AI software packages."

The AIPackage class requires the basics (name, packageVersion, suppliedBy, downloadLocation, releaseTime, primaryPurpose). It adds optional AI fields including typeOfModel, informationAboutTraining, limitation, metric, autonomyType, useSensitivePersonalInformation and safetyRiskAssessment.

One trap. The safetyRiskAssessment values (serious, high, medium, low) follow the EU general product safety risk methodology, not the EU AI Act's risk categories. Writing "high" there does not mean "high-risk" under the AI Act, and nobody reading your file should assume it does.

The DatasetPackage class requires datasetType and originatedBy, and offers dataCollectionProcess, dataPreprocessing, knownBias, hasSensitivePersonalInformation, anonymizationMethodUsed and confidentialityLevel.

Model cards

Model cards predate both formats. The idea comes from Model Cards for Model Reporting (Mitchell et al., 2018): short documents that ship with a model and describe its performance and intended use. On Hugging Face a model card is the repository's README.md with a YAML header. That header can name the license, the datasets, and a base_model for fine-tunes, adapters, quantized copies and merges.

A model card is a source for your AI-BOM, not a substitute for one. It is written by the model's publisher, it describes one model, and nothing about it ties it to the exact file you deployed. The base_model field is still the fastest way to trace a downloaded model's lineage.

The G7 minimum elements

On 12 May 2026 CISA and its G7 partners published Software Bill of Materials for AI: Minimum Elements. The document is jointly published by the cybersecurity agencies of Germany, Italy, France, Canada, the US, the UK and Japan, in collaboration with the European Commission. It says its elements are "not mandatory" and "in addition to the general SBOM minimum elements."

It sorts the elements into seven clusters: Metadata, System Level Properties, Models, Datasets Properties, Infrastructure, Security Properties, and Key Performance Indicators. The Models cluster asks for a model hash, which it defines as "the hash of the weights or model file" for each model. It also asks for lineage, including "the predecessor model on which fine tuning was done" for any fine-tune. The System Level Properties cluster asks for data flows, with examples including the APIs of external services the system uses and multi-agent communication protocols.

This is the most useful document of the four for a first inventory, because it answers "what goes in" without tying you to a format. Use it as the checklist and CycloneDX or SPDX as the container.

Where regulation asks for this

The EU AI Act never says "bill of materials." For high-risk systems it asks for the information one contains. Article 11(1) requires technical documentation that contains "at a minimum, the elements set out in Annex IV." Annex IV then asks for:

  • Point 1(c): "the versions of relevant software or firmware, and any requirements related to version updates."
  • Point 2(a): the development methods, "including, where relevant, recourse to pre-trained systems or tools provided by third parties and how those were used, integrated or modified by the provider."
  • Point 2(c): the system architecture, "explaining how software components build on or feed into each other."
  • Point 2(d): datasheets on the training data, including "information about their provenance, scope and main characteristics" and "how the data was obtained and selected."

Two more provisions make the supply chain explicit. For high-risk systems, Article 15(5) requires measures against "data poisoning" by name. It names attacks on "pre-trained components used in training (model poisoning)" in the same sentence. Article 25(4) requires the provider of a high-risk system and any third party supplying "tools, services, components, or processes" to agree in writing on the information the provider needs. Free and open-source suppliers are exempt from that, except for general-purpose AI models.

On the other side of the relationship, Annex XII lists what a general-purpose model provider must hand to companies building on its model. It includes "the licence for the model" by name. It also includes information on training, testing and validation data, "including the type and provenance of data and curation methodologies" where applicable. My piece on the AI Act for downstream providers covers how to ask for that package and when the high-risk dates now apply.

If you are working toward ISO 42001, an AI-BOM is also the plain asset inventory any management system needs underneath it. That certificate covers a process within a stated scope. The inventory is how you show what is inside the scope.

How to produce a first AI-BOM in one week

Pick one product, ideally the one with the most AI surface area. Do not try to cover the company. A complete inventory of one product teaches you more than a partial one of ten.

Day 1: find every model. Search the codebase for provider SDK clients, model ID strings, model loading calls such as from_pretrained and torch.load, and files ending in .safetensors, .bin, .pt, .pkl or .onnx. Check configuration and environment variables too, since model IDs often live there.

Day 2: describe each model. Record the producer, the exact version or dated snapshot ID, how you reach it (hosted API or weights you hold), the file format, the license, and the base model. For weights you hold, compute a SHA-256 hash of the file you actually deploy. For a hosted API you cannot hash anything. The G7 document says that when the author has no access to the artifact, the value should be marked unknown. Write "unknown" rather than leaving it blank.

Day 3: list the data. Fine-tuning sets, evaluation sets, and every retrieval corpus. For each one: where it came from, how it was collected, whether it contains personal data, who owns it, and a hash or snapshot ID where one exists.

Day 4: dependencies, services and tools. Generate a normal SBOM for the service with whatever tool you already use, then check that the AI libraries are pinned to exact versions. List every external model API and MCP server with its endpoint, version, how it authenticates, and which side of your trust boundary it sits on.

Day 5: write it down properly. Put it in CycloneDX 1.7 or SPDX 3.0.1 JSON and validate it against the schema. Add the author, a timestamp and a version, and sign it if you sign other release artifacts. Name an owner and decide what triggers an update: a new model ID, a new fine-tune, a new dataset, or a new tool.

The fields below are the minimum I would ship in that first version, mapped to where each lives in both formats and in the G7 list.

FieldWhy it mattersCycloneDX 1.7SPDX 3.0.1G7 element
Model name and exact versionAliases change underneath youname, version on a machine-learning-model componentAIPackage name, packageVersionModel name, Model version
ProducerWho to contact when it breakssuppliersuppliedByModel producer
Base model lineageA parent's flaw is inheritedpedigree.ancestorsCore RelationshipModel description (lineage)
Weights hashProves which file you ranhashesverifiedUsingModel hash value, Model hash algorithm
File formatPickle can run code on loadmime-type or propertiescomment or extensionModel properties
LicenseOpen weight is not open datalicenseshasConcludedLicense relationshipModel license
Training approach and limitationsWhat the model was never meant to domodelCard.modelParameters, considerations.technicalLimitationsinformationAboutTraining, limitationModel training properties
Datasets and provenancePoisoning enters through datadata component, modelParameters.datasetsDatasetPackage dataCollectionProcess, originatedByDataset provenance
Personal data in datasetsPrivacy and AI Act data dutiesdata.sensitiveDatahasSensitivePersonalInformationDataset sensitivity
Inference libraries, pinnedThe LiteLLM and torchtriton pathlibrary componentsSoftware PackageSystem components
Hosted APIs and MCP serversCode you call but do not holdservice with endpoints, authenticated, trustZoneNo dedicated AI classSystem data flow
BOM author, timestamp, versionA stale inventory is worse than nonemetadatacreationInfoSBOM author, SBOM timestamp, SBOM version

Where a cell says properties, comment or extension, neither format has a dedicated field and you are recording it in a general-purpose slot. That is fine for a first version. Be consistent so a script can read it later.

AI-BOM checklist

  • One named product in scope, and one named owner for its AI-BOM.
  • Every model listed by exact version or dated snapshot, never by alias.
  • Every held weights file hashed, with its format recorded.
  • No pickled model files loaded from outside your organization, or a written exception for each.
  • Base model recorded for every fine-tune, adapter and quantized copy.
  • Every dataset listed with source, collection method and personal data status.
  • AI libraries pinned to exact versions, with a normal SBOM alongside.
  • Every external model API and MCP server listed with endpoint, version and authentication.
  • Hosted API models marked "unknown" for hash, not left blank.
  • Annex XII documentation requested from any general-purpose model provider you build on.
  • Document validated against the CycloneDX 1.7 or SPDX 3.0.1 schema, dated and versioned.
  • A written list of events that trigger an update.

Frequently Asked Questions

What is an AI bill of materials?

An AI bill of materials, or AI-BOM, is an inventory of the components an AI system is built from: models and their lineage, datasets, model files, inference libraries, and the external services and tools the system calls. It extends a software bill of materials with fields that only AI components need, such as training data provenance and weights hashes.

What is the difference between an ML-BOM and an AI-BOM?

Mostly the name. ML-BOM is CycloneDX's term for its machine learning support, which covers models and datasets and first appears in the CycloneDX 1.5 schema. AI-BOM is the general term. SPDX 3.0 covers the same ground through its AI and Dataset profiles, and the G7 calls it an SBOM for AI.

Does the EU AI Act require an AI-BOM?

Not by name. For high-risk systems, Annex IV requires technical documentation covering software versions, third-party pre-trained systems and tools, system architecture, and training data provenance. That is the content of an AI-BOM. The Act does not prescribe CycloneDX, SPDX or any other format.

Should I use CycloneDX or SPDX for an AI-BOM?

Use whichever your organization already uses for SBOMs, so one toolchain handles both. CycloneDX 1.7 has a structured model card object and a service object that suits hosted APIs and MCP servers. SPDX 3.0.1 has dedicated AI and Dataset package classes with fields such as known bias and data collection process.

Are safetensors files safe to load?

Safetensors is designed to store only tensors, unlike pickle, which can execute code when a file is loaded. That removes the main code execution risk of downloaded weights. It does not tell you whether the weights themselves were trained on poisoned data, which is why lineage and dataset provenance belong in the inventory too.

How do I inventory a model I only reach through an API?

Record the provider, the exact model ID or dated snapshot, the endpoint, and how you authenticate. You cannot hash weights you do not hold, so mark the hash as unknown, as the G7 minimum elements suggest. Ask a general-purpose model provider for its Annex XII documentation to fill in license and training data fields.

The first version will be incomplete, and that is fine. The test of an AI-BOM is not whether it is perfect. It is whether, the next time a package like LiteLLM is pulled from PyPI, you can answer "are we affected" in minutes instead of days.

Get new AI Security & Agents writing

Enjoyed this? Subscribe and tell us what you read most. AI Security & Agents is already ticked for you. No tracking pixels, unsubscribe with one click.

Tell us what you read most (optional)

About DeepakPublicationsAnalysisAll tracks