AML.T0060

MITRE ATLAS / Publish Hallucinated Entities

Adversaries register an entity that a language model invents, such as a package name, website, or email address the model recommends but that does not exist, then fill it with adversary-controlled content so the model's advice leads users to them.

Think of it likeA confident tour guide keeps sending visitors to a restaurant that was never there, so you open one at that address and wait.

In plain English

Language models invent plausible package names and URLs. If a model reliably hallucinates the same one, an attacker registers it, and every developer who follows the model's advice installs their code.