Defend against adversarial examples, data poisoning, model theft, prompt injection, insecure tools, and compromised dependencies through threat modeling and testing.