AML.T0110

MITRE ATLAS / AI Agent Tool Poisoning

Adversaries poison the tools an agent uses by introducing or modifying malicious content or behavior in a tool's model-visible definition, its executable implementation, or its runtime responses.

Think of it likeRelabeling the jars in a kitchen so the recipe is followed exactly and the dish still comes out wrong.

In plain English

A tool's description is text the model reads and believes. Malicious wording in a tool definition, or a tool that returns instructions instead of data, redirects the agent while every step looks legitimate.