AML.T0024

MITRE ATLAS / Exfiltration via AI Inference API

Adversaries exfiltrate private information through legitimate inference API access, extracting training data, model parameters, or the model itself through repeated queries.

Think of it likeCopying a book by asking the librarian to read out one page at a time, day after day.

In plain English

The API is a data channel. Enough carefully chosen queries can pull back training examples, reconstruct model behavior, or recover information the model was never meant to reveal.