AML.T0031

MITRE ATLAS / Erode AI Model Integrity

Adversaries feed adversarial inputs over time to degrade a model's apparent performance, eroding user and operator confidence in the system rather than causing a single dramatic failure.

Think of it likeLoosening one screw a week until the whole team decides the machine cannot be relied on.

In plain English

This attack targets trust, not availability. A model that is quietly wrong often enough gets switched off by its own operators, which is exactly the outcome the adversary wanted.