In spite of the wide adoption of policy-based approaches for security management, and many existing treatments of pol- icy verification and analysis, relatively little attention has been paid to policy refinement: the problem of deriving lower-level, runnable policies from higher-level policies, pol- icy goals, and specifications. In this paper we present our initial ideas on this task, using and adapting concepts from data integration. We take a view of policies as governing the performance of an action on a target by a subject, possibly with certain conditions. Transformation rules are applied to these components of a policy in a structured way, in order to translate the policy into more refined terms; the transfor- mation rules we use are similar to those of ‘global-as-view’ database schema mappings, or to extensions thereof. We illustrate our ideas with an example.