AI-SBOM and ML Bill of Materials Generation with Provenance Attestation

You are the security engineer designing an AI-SBOM (machine-learning bill of materials) capability for [INSERT ORGANIZATION]'s MLOps platform, under authorization [INSERT AUTHORIZATION ID]. In-scope pipelines build and deploy [INSERT MODEL TYPES, e.g., LLMs, vision encoders, rerankers] onto [INSERT SERVING PLATFORM], using both internally trained weights and licensed third-party or hub artifacts. This is a defensive build task: do not retrieve or redistribute proprietary model weights; work only with artifacts the organization is licensed to use. Module 1 - Scope: define SBOM object granularity - base model, tokenizer, adapter/LoRA merge, dataset snapshot, embedding model, inference runtime, container base image, and attached tool/MCP descriptors. Module 2 - Schema and fields: specify supplier, version, hashes, license, training-data provenance, known-vulnerability references, safety-evaluation results, signature, and build provenance, aligned to SPDX, CISA ML BOM concepts, and SLSA-for-ML expectations. Module 3 - Generation method: describe pipeline-integrated generation at each stage - data version pinning, training-run metadata capture, reproducibility manifest, and OCI-style artifact signing with Sigstore/cosign-like tooling. Module 4 - Verification and admission: define policy that blocks unsigned or unverifiable artifacts, hash mismatch, license conflicts, and components listed in [INSERT VULNERABILITY/RECALL FEED SOURCE]. Module 5 - Detection and monitoring: propose telemetry for SBOM drift, expired or revoked attestations, new transitive dependencies, and supplier key rotation, with SIEM rule pseudocode and alert thresholds. Module 6 - Protection guidance: key custody and separation of duties, immutable transparency log, air-gapped mirror for critical base models, and patch/replacement SLAs for flagged components. Module 7 - Report: provide the SBOM JSON template, a text architecture diagram, an admission-policy example, and a gap assessment against current tooling at [INSERT CURRENT ML PLATFORM]. Include example operator queries answering "which running services contain model X version Y" and "which data snapshot was used to train artifact Z". Add a text data-flow diagram showing trust boundaries between the hub, build environment, registry, and serving environments. Map every control to clauses in [INSERT APPLICABLE FRAMEWORKS, e.g., EU AI Act GPAI obligations, NIST AI RMF, ISO/IEC 42001, Cyber Resilience Act]. State assumptions about existing CI/CD and identity infrastructure, list prerequisites, and define acceptance criteria, rollout phases, and re-attestation triggers (rebuild, adapter merge, runtime CVE, supplier advisory). Begin with the design plan and field dictionary for approval; then deliver templates and policies step by step, pausing for my review between modules. ================