CVE-2026-28500 is a security control bypass in the ONNX Python library. The 
`onnx.hub.load()`
 function accepts a 
`silent=True`
 parameter that suppresses all trust verification warnings and user confirmation prompts, allowing silent loading of models from untrusted repositories.


The design issue: ONNX Hub's integrity verification relies on a SHA256 manifest that is fetched from the same repository that hosts the models. If an attacker controls or compromises the repository, they control both the model files and the integrity manifests. There is no independent trust anchor. The 
`silent=True`
 parameter removes the last remaining safeguard: the user prompt that would normally warn "this model is from an untrusted source."


This affects ALL ONNX versions through 1.20.1. No patch is currently available. The 
`silent`
 parameter is documented and used in tutorials, so there is likely widespread usage in production pipelines and CI/CD scripts where interactive prompts are suppressed by default.


The supply chain implications are significant. ONNX is the interchange format for ML models across frameworks. Any pipeline that loads ONNX models from external repositories using 
`silent=True`
 is loading without any verification prompt, and the SHA256 check is only as trustworthy as the repository serving the manifest.