To secure an AI agent interacting with sensitive user data, I would implement data encryption both at rest and in transit, use access controls to limit who can interact with the data, and regularly audit the data access logs for any anomalies. Additionally, I would ensure the AI model is trained to avoid exposing sensitive information in its outputs.