Evaluating a security product requires the ability to conduct tests to assert that the product reacts as expected, both in terms of scalability and semantics. However, the production of evaluation data at a large scale with a high semantic is very costly with current methods. Load tests are semantically poor and semantic tests require a testbed environment to be deployed at a large scale. Evaluation data from real world activity need to be anonymized and a compromise must be made between the request of the evaluator and the interest of the real world organization. Moreover, to evaluate the full scope of a security product, the evaluator needs multiple test methods. In this paper, we describe a new methodology to produce evaluation data with a customizable level of realism and the possibility to be deployed at a large scale with lower resource requirements for a network support than a testbed environment. Our prototype relies on this method to generate realistic activity for up to 250 simulated users interacting with a real-world webmail server.