GuideAIBOMCERT-In v2.0 · §9

What is an AIBOM?

CERT-In's Technical Guidelines v2.0 (July 2025) added the AIBOM to the set. Here is what section 9 asks for, element by element, how each element lands in a CycloneDX document, and a schema-valid sample you can download.

The definition, in CERT-In's words

Section 9.1 defines an AIBOM as “a comprehensive list of components used in building, training, and deploying AI models.” It typically includes hardware (servers, sensors, GPUs), software (AI models, frameworks, development tools), data sources, and anything else an AI implementation needs.

The guidelines give four reasons for it (§9.2). Security: a breakdown of the system shows its weak spots, and speeds up finding compromised components during an incident. Transparency: stakeholders can see how a model was built and what data trained it. Compliance: with regulation, industry standards and ethical guidelines, including data-protection rules. Risk management: dependencies and failure points become visible.

The 19 minimum elements, at a glance

Table 10 lists its elements in one column. Grouped by the question each one answers, they read like this — select any element to open it in the explorer below. (The grouping is ours, not CERT-In's.)

Identity

How it is built

Data

Behaviour

Use

Risk and trust

Figure 1 · Table 10's 19 elements, grouped by KavachQ for reading; numbers are the element's position in Table 10. Source: CERT-In Technical Guidelines v2.0, Table 10, pp. 54–55.

Element by element

Pick an element to see what CERT-In asks for, an illustrative value for a document-classification model, and where it sits in a CycloneDX document. CycloneDX describes a model as a machine-learning-model component carrying a model card, which covers many of Table 10's elements natively. The sample on the right is a real, schema-valid CycloneDX 1.6 file — the highlighted lines change as you go.

Turn on JavaScript to explore the elements, or read the full list in Table 10 of the guidelines.


        
aibom-example.cdx.json · CycloneDX 1.6 · validated against the official schemaDownload the sample
Figure 2 · Table 10 elements with an illustrative document classifier. The model, developer, libraries, figures and advisory are fictional, for teaching only. Elements without a native CycloneDX field are carried as named properties. Source: CERT-In Technical Guidelines v2.0, Table 10, pp. 54–55.
WHERE AI MEETS QUANTUM RISK

Two of Table 10's elements are cryptography: Security Requirements (the encryption and access control protecting model data) and Attestations (a digital signature for the model or AIBOM). A model or AIBOM signed with RSA or ECDSA carries a signature that a large enough quantum computer could forge. Recording the signature algorithm lets you find those signatures and plan to move them to a post-quantum scheme such as ML-DSA (FIPS 204).

What §9.4 asks organisations to do

CERT-In's AIBOM recommendations, in brief. Read the clause text in the guidelines for the exact wording.

  1. §9.4.1.1Government, public-sector and essential-services organisations, and organisations in AI-driven development and services, should include AIBOM requirements in all AI-related procurements and solutions.
  2. §9.4.1.2All AI solutions supplied to government, public-sector and essential-services organisations must be accompanied by an AIBOM.
  3. §9.4.1.3Those organisations must maintain an AIBOM for every AI system they use, procure or develop.
  4. §9.4.1.4AIBOMs supplied to government and the public sector must include the elements in the guidelines.
  5. §9.4.1.5Use established standards such as SPDX or CycloneDX.
  6. §9.4.1.6After a vulnerability or security issue is found, the AI supplier should issue a VEX for the AI model: Not Affected, Affected, Fixed or Under Investigation.
  7. §9.4.1.7Integrate AIBOM data with vulnerability databases, CERT-In vulnerability notes and alerts, threat intelligence and vendor advisories.
  8. §9.4.1.8Consumer organisations build an internal AIBOM from the supplier's.
  9. §9.4.1.9Security teams bring the AIBOM into vulnerability-management workflows.
  10. §9.4.1.10Audit the AIBOM process regularly for accuracy, completeness and compliance.
  11. §9.4.1.11Combine AIBOM component data with VEX status for an up-to-date view of the AI system's vulnerabilities.
  12. §9.4.1.12Store and transmit AIBOM data securely — encryption, access control and other measures.
  13. §9.4.1.13Update the AIBOM as new AI components or models are introduced or existing ones are updated.

The best practices in §9.4.2: keep a comprehensive inventory of the system's critical components; use a standard format such as CycloneDX or SPDX; automate AIBOM generation in model development and deployment pipelines; include scripts, model weights and configuration so results can be reproduced; start with high-priority or high-risk models; and track model lineage — versions, retraining, and modifications and enhancements.

CERT-In's guidelines use “should” and “must”, but they are technical guidelines; how they bind you depends on your regulator and your contracts.

How KavachQ produces an AIBOM

KavachQ generates the AIBOM as its own CycloneDX document against Table 10, and every entry says how KavachQ knows it:

  • Observed entries come from the scan: AI frameworks and model-provider SDKs found in the SBOM, and, from a code scan, model files and the places your code calls a model API. A framework being present is not evidence that a model is deployed, and the entry says so.
  • Declared entries come from your AI-model registry. Training data, performance metrics and intended and out-of-scope usage can only come from whoever built the model, so KavachQ carries your statement and labels it as declared.
  • Where a model file has a signature beside it, the AIBOM records the signature algorithm and whether the signer is quantum-vulnerable.
  • Every Table 10 element is present — on every component, or for Attestations on the document — either as a value or explicitly marked unknown, not applicable or withheld; never silently missing.
  • The signed variant carries an ML-DSA (FIPS 204) signature over the document, verifiable offline against KavachQ's published trust anchor — Table 10's Attestations element.

AIBOM: frequently asked questions

What is an AIBOM?

In CERT-In's Technical Guidelines v2.0 (9 July 2025), an Artificial Intelligence Bill of Materials (AIBOM) is a comprehensive list of the components used in building, training and deploying AI models. It typically includes hardware such as servers, sensors and GPUs; software such as AI models, frameworks and development tools; data sources; and any other elements needed for AI implementation. Table 10 sets 19 minimum elements.

What are the minimum elements of an AIBOM?

Table 10 lists 19: model name, model version, model type, model developer, model licensing information, software dependencies, ML models and algorithms, model performance metrics, data source, data sets, hardware, security requirements, input, output, intended usage, out-of-scope usage, environmental impact, vulnerabilities and attestations — a digital signature for the model or AIBOM.

Who has to provide an AIBOM?

Section 9.4.1.2 says all AI solutions supplied to government, public-sector and essential-services organisations must be accompanied by an AIBOM. Section 9.4.1.1 asks those organisations, and organisations involved in AI-driven development and services, to include AIBOM requirements in all AI-related procurements and solutions.

What format should an AIBOM use?

Section 9.4.1.5 says it should follow established standards such as SPDX or CycloneDX. CycloneDX describes a model through a machine-learning-model component with a model card; elements it has no field for, such as out-of-scope usage or hardware, can be carried as named properties.

Does an AIBOM have to be signed?

Table 10 lists Attestations — a digital signature for the model or AIBOM to ensure authenticity and integrity — among its minimum elements.

How is an AIBOM different from a model card?

A model card describes one model: what it does, how it performs and what it should not be used for. An AIBOM inventories the whole AI system around it: the models, the software they depend on, the training data, the hardware, known vulnerabilities and a signature. In CycloneDX, the model card sits inside the AIBOM, on the model's component.

Sources

  1. CERT-In (Ministry of Electronics and Information Technology), Technical Guidelines on SBOM | QBOM & CBOM | AIBOM | HBOM, Version 2.0, 9 July 2025 — §9 (pp. 53–58), Table 10 (pp. 54–55).
  2. OWASP CycloneDX, CycloneDX 1.6 JSON schema — modelCard, componentData.
  3. NIST, FIPS 204, Module-Lattice-Based Digital Signature Standard (ML-DSA), August 2024.

Related: xBOM — all five bills of materials · What is an SBOM? · What is an HBOM? · What is a QBOM? · What a CBOM actually is · KavachQ's AI policy.

AIBOM, signed

Know what your AI runs on — and how it is signed

Start with a free scan of a public domain, then scope a pilot that produces a signed AIBOM and CBOM for a real slice of your estate.