← AI security in plain English

AML.TA0000

MITRE ATLAS / AI Model Access

The adversary seeks some level of access to the AI model itself, using techniques such as AI Model Inference API Access, AI-Enabled Product or Service, and Full AI Model Access.

Think of it likeGetting close to the safe: peering through the keyhole, using it as a paying customer, or holding the full combination and its internal workings.

In plain English

Attackers get their hands on the model, from simply querying its API to fully accessing its internals, so they can probe it, feed it data, and craft attacks.