AML.T0063

MITRE ATLAS / Discover AI Model Outputs

Adversaries obtain model outputs that are not required for the system to function and are not intended for end users, such as class confidence scores, logits, or embedding vectors.

Think of it likeBeing handed the examiner's full mark sheet when all you needed was the pass or fail.

In plain English

Confidence scores and raw logits leak far more than a label does. Returning them makes model extraction and evasion much cheaper for anyone querying the API.