Marty Bent says AI agents should be secured using adversarial assumptions, not trust, drawing on Bitcoin’s resilience model. He argues robust systems need hard choke points and controls instead of relying on constant alignment.