A security checklist for AI in production

Before an agent touches production systems, make sure every permission is scoped, every action is logged and every sensitive step needs approval.