AML.T0080

MITRE ATLAS / AI Agent Context Poisoning

Adversaries manipulate the context an agent's language model works from, influencing the responses it generates and the actions it takes. Memory files, scratchpads, prior turns, and retrieved documents are all context.

Think of it likeEditing the notes someone left themselves overnight, so the plan they wake up and follow is not the one they wrote.

In plain English

An agent's context is its working memory. Anything an attacker can write into that memory shapes every decision the agent makes afterward, long after the original injection.