The important aspect of performance of a multi-agent system deployed in highly non-homogeneous environment is ensuring data confidentiality. In such an environment each agent resides in its local ecosystem characterized by its own data structure and semantics. Moreover agents may also extend their knowledge by sending appropriate queries to other agents. Thus a knowledge diffuses through the system so it may cause unauthorized data accesses. In this article we propose the run-time method of discovering such data leaks and the method of preventing such events.