MITRE ATLAS / Persistence
The adversary tries to keep their foothold across restarts and cleanups, using techniques such as Poison Training Data, Manipulate AI Model, and RAG Poisoning.
Attackers plant lasting backdoors in the AI: poisoned training data, a tampered model, or corrupted knowledge sources, so they keep access after reboots and fixes.