AML.T0057

MITRE ATLAS / LLM Data Leakage

Adversaries craft prompts that induce a model to reveal sensitive information from its training data, its context window, or the systems it can reach.

Think of it likeSteering a conversation until someone discreet repeats something they were told in confidence.

In plain English

A model does not know which parts of its knowledge are secret. The right question can surface training data, another user's context, or internal documents it retrieved.