The design principle of restricting local autonomy only where necessary for global robustness has led to a scalable Internet. Unfortunately, this scalability and capacity for distributed control has not been achieved in the mechanisms for specifying and enforcing security policies. This shortcoming must be overcome if end-to-end security mechanisms (such as IPsec or TLS) are to ever replace solutions of short-term convenience such as firewalls. The STRONGMAN (for Scalable TRust Of Next Generation MANagement) system offers three new approaches to scalability, applying the principle of local policy enforcement complying with global security policies. First is the use of a compliance checker to provide great local autonomy within the constraints of a global security policy. Second is a mechanism to compose policy rules into a coherent enforceable set, e.g., at the boundaries of two locally autonomous application domains. Third is the "lazy instantiation" of policies to reduce the amount of state that enforcement points need to maintain. We demonstrate the use of these approaches in the design, implementation, and measurements of a distributed firewall. Our experiments show that, under certain circumstances, performance can improve over the traditional-firewall approach.