← AI security in plain English

AML.T0040

MITRE ATLAS / AI Model Inference API Access

Adversaries obtain access to a model through its legitimate inference API. Query access alone is enough to support extraction, evasion development, and probing for training data.

Think of it likeLearning how a vending machine works by feeding it coins all day, never once opening the cabinet.

In plain English

You do not need the weights to attack a model. Enough queries against a public API reveal its decision boundaries, and sometimes enough to rebuild a usable copy of it.