Authoritative Guide to AI/ML-BOM

⚡ Quick Summary

The Authoritative Guide to AI/ML-BOM, developed by the OWASP CycloneDX AI/ML Working Group and ratified under Ecma International (ECMA-424), provides a comprehensive standard for Machine Learning Bills of Materials (ML-BOMs). As artificial intelligence systems grow in complexity, software supply chain transparency must extend beyond traditional software dependencies to include model metadata, neural network architectures, training datasets, hardware and container runtime topologies, and tokenizers. The guide outlines normative data structures and best practices using CycloneDX v1.7 to document model lineage, performance benchmarks, ethical limitations, fairness assessments, and energy consumption metrics. It delivers a standardized foundation for managing AI supply chain security, tracking data provenance, verifying model reproducibility, and fulfilling regulatory obligations.

🧩 What's Covered

The specification delivers an end-to-end blueprint for declaring and managing machine learning components within the CycloneDX BOM schema:

• Declaring ML Models & Repositories: Modeling ML repositories as components and assemblies, capturing package URLs (PURLs), commit hashes, custom or SPDX licensing, release notes, and hierarchical lineage graphs covering ancestors and descendants across fine-tuning, quantization, format conversions, pruning, and adapter modifications.
• Model Parameters & Architecture: Encoding learning approaches (supervised, unsupervised, reinforcement, semi-supervised, self-supervised), ML tasks, neural network architecture families (Transformers, CNN, RNN, LSTM, GRU, GAN), framework configurations, and hyperparameters via the cdx:ai-ml:model property taxonomy.
• Datasets & Data Lineage: Managing public and private datasets as in-line descriptors or standalone data component references, detailing data curation methods, acquisition mechanisms, classifications, and access control boundaries.
• Quantitative Analysis & Benchmarks: Documenting standardized benchmark evaluations (such as MMLU, GLUE, GSM8K, HumanEval, ImageNet), raw performance metrics with statistical confidence intervals, and base64-encoded comparative evaluation graphics.
• Considerations & Responsible AI: Capturing intended user personas, operational design domain use cases, technical limitations (including hallucination, context boundaries, and synthetic data degradation), performance tradeoffs, ethical risks with mitigation strategies, and demographic fairness assessments.
• Environmental Impact & Manufacturing: Quantifying lifecycle energy consumption per activity (design, training, validation, inference) in kWh, tracking energy source mixes, and recording metric tonnes of CO2 equivalent (tCO2eq) costs and offsets.
• EU AI Act Compliance Mappings: Detailed mappings aligning CycloneDX schema elements with Article 53 GPAI obligations, Annex XI technical documentation, and the European Commission's mandatory Public Summary of Training Content template.

💡 Why it matters?

As regulatory scrutiny accelerates globally—most notably through the European Union's Cyber Resilience Act (EU CRA) and the EU AI Act—organizations must transition away from informal, unstructured model cards and research papers toward machine-readable, auditable standards. The CycloneDX ML-BOM standard enables automated vulnerability tracking, data provenance verification, and copyright compliance for text and data mining (TDM). It gives governance and engineering teams an interoperable framework to ensure reproducibility, audit model risks, and demonstrate conformity to regulators and downstream system integrators.

❓ What's Missing

While the guide provides robust mappings for CycloneDX v1.7, it acknowledges that certain areas remain emerging. Formal schema objects for configuration parameters, hyperparameters, and domain-specific dataset acquisition workflows currently rely on generic property taxonomy extensions rather than fully dedicated top-level schema types. Additionally, normative definitions for AI/ML workflow task types and fine-tuning taxonomies are slated for CycloneDX v2.0, meaning current implementations require custom conventions for deep workflow orchestration and real-time automated crawlers.

👥 Best For

AI governance professionals, MLOps engineers, software supply chain security specialists, compliance officers, enterprise risk architects, and organizations building or deploying general-purpose AI (GPAI) models subject to regulatory documentation mandates.

📄 Source Details

• Title: Authoritative Guide to AI/ML-BOM: Drive Transparency, Compliance, and Security Across the AI Supply Chain
• Publisher: OWASP Foundation & OWASP CycloneDX ( ratified as ECMA-424 by Ecma International)
• Edition / Date: First Edition (Revision 1), 10 June 2026 (Initial Release: 3 March 2026)
• License: Creative Commons Attribution 4.0 International (CC BY 4.0)

📝 Thanks to

Developed by the CycloneDX AI/ML Working Group. Key primary contributors include Matt Rutkowski (IBM, Chair AI/ML Work Group), Steve Springett (Chair of CycloneDX Standard, Founder Ecma TC54, Chair of OWASP Global Board of Directors), Jan Kowalleck (CycloneDX Project Co-Lead), Michael Boone (NVIDIA), Jessica Butler (NVIDIA), Pratyusha Maiti (NVIDIA), Saquib Saifee (IBM), Colin Gigool (IBM), Aliza Heching (IBM), Nagalakshmi Satyanarayanan (IBM), and Pavel Shukhman (Reliza), with preface by Andrew van der Stock (Executive Director, OWASP Foundation).