AI agents need a security model that assumes things will go wrong. The right response isn't better permission checks, it's architecture that makes trust unnecessary.