AML.T0029

MITRE ATLAS / Denial of AI Service

Adversaries flood an AI-enabled system with requests to degrade or shut down the service, exploiting the fact that inference is far more expensive per request than serving a static page.

Think of it likeFilling a small restaurant with people who order nothing, so no real customer can get a table.

In plain English

Inference is expensive. It takes far fewer requests to exhaust a model-serving system than a conventional web service, so the bar for a successful denial of service is low.