How do you prevent prompt injection attacks in your AI features?

Describe the layered defenses that keep untrusted input from steering your AI features, such as separating system instructions from user-supplied content, constraining model outputs to structured formats, and validating any action a model proposes before it executes. Confirm that AI components run with least-privilege access so a successful injection cannot reach data the requesting user could not already see, and that you test these defenses through adversarial red-teaming.