In the world of AI security, we have long focused on software-level threats: data poisoning, prompt injection, and API abuse.…