A New Security Frontier
The Open Source Security Foundation (OpenSSF) published its AI Supply Chain Security Framework this week, offering the first industry-consensus guidance on securing AI model supply chains in enterprise software. The framework addresses a gap identified by security researchers: while software supply chain security has matured significantly since the SolarWinds attack in 2020, AI models integrated into products represent an analogous but distinct risk category that existing frameworks do not fully cover.
The Threat Model
AI supply chain risks include: adversarial model poisoning (manipulating training data to cause model misbehavior), model substitution (replacing a legitimate model with a compromised version), prompt injection via training data, and unintended capability acquisition (models learning behaviors not intended by developers). These risks are distinct from traditional software vulnerabilities and require tailored mitigations.
Framework Components
The OpenSSF framework covers four domains: model provenance verification (cryptographic attestation of model lineage), training data auditing (documentation and verification of training data sources), behavioral testing (standardized tests for unintended capabilities before production deployment), and runtime monitoring (ongoing checks for unexpected model behavior in production).
Tool Support
The framework is accompanied by open-source tooling: model transparency reports (analogous to SBOM for software), provenance signing using Sigstore integration, and a model evaluation harness compatible with common deployment environments.
Adoption
The Linux Foundation, Microsoft, Google, and Amazon have all endorsed the framework. Several AI deployment platforms including Databricks and Hugging Face have announced planned support for framework-compatible model attestation.