← AI security in plain English

AML.T0092

MITRE ATLAS / Manipulate User LLM Chat History

Adversaries alter a user's chat history with a language model to conceal malicious activity, removing or rewriting the turns that would reveal what was asked and what the model did.

Think of it likeTearing the incriminating pages out of the visitor log before anyone reviews it.

In plain English

Chat history is often the only record of what an assistant was told to do. Editing it removes the evidence and leaves the user believing the conversation went differently.