ONNX Runtime parser backdoor lets models exfiltrate data
Researchers discovered an ONNX Runtime parser vulnerability that allows malicious neural network models to execute arbitrary file and memory reads upon loading. The authors demonstrate how this flaw can be weaponized in open-weight language models like Qwen2 to quietly exfiltrate sensitive server data when triggered by a specific prompt. The findings highlight the security risks of deploying unvetted AI models from public repositories, arguing that they should be scrutinized as carefully as untrusted source code. Open article →