← AI security in plain English

AML.T0076

MITRE ATLAS / Corrupt AI Model

An adversary deliberately corrupts a malicious model file so it cannot be fully deserialized, causing a model scanner to fail or skip the file while the target's loader still reaches the payload.

Think of it likeDamaging a parcel just enough that the scanner rejects it and a human waves it through by hand.

In plain English

Scanners have to parse a file to inspect it. Breaking the format in a way that defeats the scanner but not the loader is a way to be neither read nor blocked.