Model-risk oversight is the governance used to test, approve, monitor, and constrain AI or statistical models before they influence business decisions. In regulated environments, it covers data quality, validation, drift monitoring, access control, human review, documentation, and rollback procedures. The goal is to make model behavior explainable, repeatable, and auditable.
In cyber security, this matters because a model can become part of the attack surface. Poisoned training data, prompt injection, overly broad retrieval, or silent model drift can produce unsafe outputs, leak sensitive information, or automate bad decisions. Strong oversight helps defenders detect abuse, require change control, and keep AI-assisted workflows bounded. It is a control-layer defense: the model may assist, but people and policy still make the final call where risk is material.