Infrastructure as code (IaC) for the cloud, which automatically configures a system’s cloud environment from source code, is an important practice thanks to its efficient, reproducible provisioning. On a cloud IaC definition (template), developers must carefully manage permission settings to minimize the risk of cyber-attacks. To this end, least privilege on IaC templates, i.e., the assignment of a necessary and sufficient set of permissions, is widely regarded as a best practice. However, the discovery of least privilege can be an error-prone, burdensome task for developers. This is partially because the execution of an action on the cloud sometimes implicitly requires permissions of other services, and since these are difficult to recognize without actual execution, developers are forced to manually iterate the execution of an action and the modification of permissions. In this work, we present an approach to automatically discover least privilege. Our approach utilizes a test suite, which represents what a system should achieve on the cloud, as an indicator of least privilege, and it iterates testing on the cloud and (re)configuration of permissions on the basis of the test results. We also propose a stepwise filtering technique that utilizes the co-occurrences of cloud services/actions and clustering-based pruning to efficiently rule out unnecessary permissions. Our experiments demonstrate that this filtering reduces the number of iterations compared to naive approaches, which directly affects the time and cost to discover least privilege. Moreover, three case studies show that our approach can identify least privilege on Amazon Web Services within a practical time.