LotL Attack Hides Malware in Windows Native AI Stack

According to Dark Reading, a researcher has demonstrated that Windows’ native artificial intelligence stack can act as a vector for malware delivery, revealing a living-off-the-land (LotL) technique that utilises trusted ONNX files. The proof-of-concept, shared exclusively with Dark Reading, shows how malware could be embedded in ONNX models and loaded by Windows’ AI inference pathway through the Windows Machine Learning API, leveraging signed Microsoft DLLs to evade detection.

The attacker could transmit such a malicious ONNX file via phishing emails or by publishing a compromised model on an open-source platform like Hugging Face, exploiting the widespread trust in AI components. The piece notes that security tools may fail to flag this because ONNX files are treated as benign AI data during loading, challenging traditional EDR and antivirus monitoring.

It also suggests mitigations, including reworking detection to monitor who loads ONNX models, what data is extracted, where that data is passed, and applying static analysis with tools like YARA, alongside AppLocker.