AI models that learn from live data streams continuously update their parameters based on incoming data. This creates two specific risks: the model ' s behavior may drift from its validated state as data patterns change (data drift), and adversaries may deliberately introduce malicious data to manipulate the model ' s learning (data poisoning).

Why D is Correct: According to ISACA AAIR adaptive model risk guidance, implementing automated monitoring for both data drift and data poisoning is the most comprehensive response to live-learning model risks. Automated monitoring operates continuously at the speed of the data stream, detecting statistical changes in input distributions (drift signals) and anomalous data patterns (poisoning signals) in real time—enabling timely intervention before either risk materializes into harmful behavior.

Why A is Wrong: Defense-in-depth for model access controls who can interact with the model but does not address risks arising from the data the model learns from. Access controls are necessary but insufficient for managing adaptive learning risks.

Why B is Wrong: Restricting data sources reduces learning breadth, potentially undermining the model ' s adaptive capability that creates its value. Periodic inspections are too infrequent for live-learning systems where risks can emerge between inspection cycles.

Why C is Wrong: Dynamic performance thresholds detect output degradation after drift has occurred. While useful as a safety net, this reactive monitoring does not prevent drift or detect poisoning early enough for the live-learning risk context.