Which AIBOM Fields Matter for Your Role?
An AI Bill of Materials (AIBOM) can carry dozens of fields. Nobody needs all of them. Answer 7 questions about your job and get the six fields that matter for your role, why each one matters to you, where it lives in CycloneDX, and the order to read the rest of the AIBOM Library.
Who this is for: CISOs and security leaders, SOC analysts, GRC and compliance teams, and AppSec and AI-security engineers who have been handed an AIBOM, or asked to produce one, and want to know where to start.
Seven questions about your job
Pick the answer closest to your week, not your job title. If two fit, pick the one you would do first.
1. When a new AI model shows up in your organisation, what is your part?
2. Which question lands on your desk most often?
3. Who do you report AI risk to?
4. Which document do you open first?
5. Which of these would hurt you most?
6. Which names do you deal with most?
7. In one line, what do you need from an AIBOM?
Answer all seven to see your role, your six fields and your reading path.
What this tool is
- A 7-question quiz, about two minutes, on what you actually do with AI models at work.
- It places you in one of four roles: CISO or security leader, SOC analyst, GRC and compliance, or AppSec and AI-security engineer.
- You get the six AIBOM fields that matter most for that role, why each one matters to you, where it lives in CycloneDX, and the value from a fictional worked example.
- You also get a reading path through the AIBOM Library, in order, built from the pages that are live today.
- Free, no sign-in. Runs in your browser; your answers are not sent anywhere.
The six fields for each role
The quiz sorts you into one of these four. Field names follow the AIBOM Format Field Map, and the example values come from Northwind Health’s Triage Assistant v2.3, a fictional company and model used across the AIBOM Library. For the full list of fields, see the AIBOM parameters checklist.
CISO / Security Leader
Decide which AI models and vendors are acceptable, and report AI supply-chain risk to the board.
| Field | Why it matters to you | CycloneDX |
|---|---|---|
| Provider, supplier and authors | The name on the contract and the party you escalate to. No named provider means no one to hold to account. | component.supplier, manufacturer, authors |
| Model licence | A fine-tune inherits obligations from its base model. Check both before legal signs. | component.licenses[] (SPDX id, name or expression) |
| Dataset licence and use conditions | Training data you had no right to use becomes your problem once the model is in your product. | licenses[] on the "data" component |
| Out-of-scope and prohibited uses | The line the vendor drew. If your use case is on the wrong side of it, the approval is yours to defend. | No field; considerations.technicalLimitations[] as a workaround |
| Signature and authenticity | Whether anyone vouches for the weights you received, cryptographically, rather than in a slide. | signature (JSF) signs the BOM data, not the weights; link a model signature via externalReferences[] type "digital-signature" |
| Safety risk level and risk management | What risk process the provider ran and where the result is. Ask for the document, not the adjective. | externalReferences[] type "risk-assessment" (a link, not a level) |
The quiz result also shows each field in the Hugging Face model card and the Northwind value. Reading path for this role.
SOC / Blue-team Analyst
Know exactly which model, version and components were running when something goes wrong.
| Field | Why it matters to you | CycloneDX |
|---|---|---|
| Model version | The first question in any AI incident: which release was live when it happened. | component.version |
| File hashes | Proves the file on the server is the file in the BOM. A mismatch is a finding on its own. | component.hashes[] (alg, content) |
| Serving stack and software dependencies | When an advisory lands for the serving framework, this is how you know whether you are affected. | components[] (framework, library, application) + dependencies[] |
| External services, APIs and tool servers (MCP) | What the model can call and change, for example a tool server with write access. That is your blast radius. | services[] (endpoints, authenticated, trustZone, data flows) |
| Retrieval index (RAG) data | Content pulled in at run time is a path for prompt injection. Know which index and when it last refreshed. | A "data" component plus services[].data; no RAG-specific field |
| Change history between versions | What changed between the version that behaved and the one that did not. | component.releaseNotes (type, title, timestamp, resolves, notes) |
The quiz result also shows each field in the SPDX 3.0 and the Northwind value. Reading path for this role.
GRC / Compliance
Turn AI documentation into evidence for the EU AI Act, ISO/IEC 42001 and NIST AI RMF.
| Field | Why it matters to you | CycloneDX |
|---|---|---|
| Intended use | Everything else is judged against it. The EU AI Act Annex IV starts with the intended purpose. | modelCard.considerations.useCases[] |
| Training datasets used | Annex IV asks for the training data sets used. A name and version is evidence; "internal data" is not. | modelCard.modelParameters.datasets[] (inline or ref to a "data" component) |
| Personal and sensitive data | Where AI documentation meets data protection: was personal data in it, and what was done about it. | data[].sensitiveData[], data[].classification |
| Performance metrics and results | Accuracy claims need a metric, a value and a named test set, or they cannot be checked. | modelCard.quantitativeAnalysis.performanceMetrics[] (type, value, confidenceInterval) |
| Human oversight and autonomy | A required item in Annex IV for high-risk systems, and the first thing a reviewer looks for in a decision-support tool. | No field |
| Standards applied and conformity | Which standards are claimed, and where the conformity evidence lives. | definitions.standards + declarations (claims, attestations, affirmation) |
The quiz result also shows each field in the EU AI Act Annex IV and the Northwind value. Reading path for this role.
AppSec & AI-Security Engineer
Generate, validate, sign and keep AIBOMs current in the build pipeline.
| Field | Why it matters to you | CycloneDX |
|---|---|---|
| Package URL or persistent ID | A package URL lets scanners and other BOMs reference the model without guessing. | component.purl, bom-ref (also omniborId, swhid) |
| Base model (fine-tuned from) | Your model inherits the base model's weights, licence and known issues. Record the exact id and revision. | component.pedigree.ancestors[] |
| Model file format (serialization) | Pickle-based formats can run code on load; safetensors cannot. CycloneDX and SPDX have no dedicated field for it, so record it on purpose. | component.mime-type or a properties[] entry; no dedicated field |
| File hashes | One hash per weight file, generated in the pipeline, so drift and tampering show up in a diff. | component.hashes[] (alg, content) |
| Signature and authenticity | A model signature from CI, linked from the BOM, is what lets the next team trust your artefact. | signature (JSF) signs the BOM data, not the weights; link a model signature via externalReferences[] type "digital-signature" |
| Build or training provenance record | Which pipeline produced the artefact from which inputs. This is where SLSA-style attestations attach. | formulation[] (workflows, tasks); externalReferences[] type "attestation" |
The quiz result also shows each field in the SPDX 3.0 and the Northwind value. Reading path for this role.
How the quiz decides
Each answer adds points to one role, sometimes a smaller number to a second. The role with the most points wins; a tie goes to your answer to the first question. If another role scores at least 60% of the winner, the result names it so you can read that path too. The questions:
- When a new AI model shows up in your organisation, what is your part?
- Which question lands on your desk most often?
- Who do you report AI risk to?
- Which document do you open first?
- Which of these would hurt you most?
- Which names do you deal with most?
- In one line, what do you need from an AIBOM?
Frequently asked questions
Which AIBOM fields should a CISO care about?
The ones that support an approve-or-block decision: the provider, the model licence and the training-data licence, the out-of-scope uses, whether the weights are signed, and where the provider's risk assessment is. A CISO rarely needs hyperparameters or architecture details.
How does a SOC analyst use an AIBOM?
As an inventory to query during an incident. The fields that matter are the model version, the file hashes, the serving stack (to match advisories), the external services and tool servers the model can call, the retrieval index it reads, and the change history between versions.
Which AIBOM fields map to the EU AI Act?
Annex IV of Regulation (EU) 2024/1689 asks high-risk system providers for, among other things, the intended purpose, the training data sets used, the metrics used to measure accuracy and robustness, the human oversight measures, and the harmonised standards applied. Each of those has a matching AIBOM field. The Field Map lists the item numbers.
What does an engineer need in an AIBOM that others do not?
Machine-checkable fields: a package URL or other persistent identifier, the base model with its exact revision, the weight file format, a hash per file, a signature, and a build provenance record. Those are what let a pipeline generate, compare and verify the file without a human reading it.
Is this quiz scored or certified?
No. It is a sorting aid. Each answer adds points to one or two roles, and the role with the most points decides your fields and reading path. If a second role scores close, the result shows it too.
More free security tools
The tools people use most on PlayCISO.
Catch injected instructions hiding in content your AI agent should only read.
Score your readiness against the first international standard for AI security threats.
Critical CVEs across 21 major vendors, month by month — and what the 2026 surge means.
Build an ordered ruleset, replay labelled packets and catch shadowed and any-to-any rules.
Self-assess all six Functions, from Govern to Recover, and see your weakest one.
Search 172 threat actors across Microsoft, CrowdStrike, Mandiant and community names.