Abstract. Whenever agents deal with condential information, it is important that they comply with a principled security policy. We show how the database concept of multi-level security can be applied to inter-agent communication. This includes the case where an unauthorized agent is misinformed on purpose in order to protect condential information.