← AI security in plain English

AML.T0062

MITRE ATLAS / Discover LLM Hallucinations

Adversaries prompt language models repeatedly to identify entities the model invents, such as nonexistent packages or URLs, that could later be registered and controlled.

Think of it likeAsking for directions over and over until you notice everyone is sent to the same street that does not exist.

In plain English

Models hallucinate consistently, not randomly. Finding a package name a model reliably recommends but that nobody has registered is the setup for a supply chain attack.