AAISM classifies model inversion as a privacy/information-leakage threat where adversaries infer or reconstruct sensitive training data or attributes from model outputs—directly jeopardizing confidential information targeted by APTs. While data poisoning, unauthorized tuning, and model distillation present material risks (integrity, governance/IP theft), the scenario’s stated objective—accessing confidential information—most directly maps to inversion. Accordingly, AAISM prioritizes defenses such as output regularization, confidence suppression/calibration, overfitting controls, privacy-preserving techniques, and strict access/telemetry on inference interfaces.