An attack where adversaries reconstruct a model's private training data by exploiting its confidence scores and outputs.
If an LLM is trained on sensitive enterprise IP, adversaries can prompt it systematically to reproduce parts of the original training documents. Defending against Model Inversion requires Output Sanitization, Differential Privacy during training, and secure enterprise boundaries that keep unauthorized endpoints from reaching the model.
See also Differential Privacy, Secure Enterprise Boundaries, Air-Gapped Environment, Role-Based Access Control and Zero Trust Architecture.