Proposal: AI Behavioral Risk Assertions Across SBOM/VEX Ecosystem (AIVEX Concept)

[devashridatta-dotcom]

Summary

Propose a standardized model for AI Behavioral Risk Assertions that extends existing SBOM/VEX concepts to AI/ML systems, enabling consistent representation of model behavior risks, prompt injection susceptibility, data provenance risks, and runtime behavioral constraints.

Motivation

Current SBOM/VEX standards focus on software component vulnerabilities but do not capture AI-specific behavioral risks. As AI becomes embedded in software supply chains, there is a need to represent non-deterministic and behavior-driven risks.

Proposal

Introduce a lightweight assertion layer (AIVEX-like model) that can:

• Represent AI model behavioral risks (hallucination, prompt injection susceptibility, unsafe outputs)
• Express provenance of training and fine-tuning data
• Link AI components to SBOM entries (CycloneDX/SPDX)
• Provide machine-readable assertions compatible with OpenVEX-style structures

Relationship to Existing Work

• CycloneDX SBOM extensions (component-level metadata)
• OpenVEX (vulnerability/exemption semantics)
• SPDX AI/ML profile discussions (where applicable)

Request

Feedback on:

1. Whether this belongs in SBOM Everywhere SIG or a new working group
2. Alignment with existing OpenSSF AI security initiatives
3. Interest in prototyping a reference schema