Review the training / MLOps pipeline for security weaknesses — data-poisoning surface, feature-store and data-source trust, experiment-tracking and model-registry access control, secrets, and reproducibility/provenance. Use when assessing how models are built and promoted, not how they're served or consumed.
