Atypical behaviours are the basis of a valuable knowledge in domains related to security (e.g. fraud detection for credit card [1], cyber security [4] or safety of critical systems [6]). Atypicity generally depends on the isolation level of a (set of) records, compared to the dataset. One possible method for finding atypic records aims to perform two steps. The first step is a clustering (grouping the records by similarity) and the second step is the identification of clusters that do not correspond to a satisfying number of records. The main problem is to adjust the method and find the good level of atypicity. This issue is even more important in the domain of data streams, where a decision has to be taken in a very short time and the end-user does not want to try several settings. In this paper, we propose Mrab, a self-adjusting approach intending to automatically discover atypical behaviours (in the results of a clustering algorithm) without any parameter. We provide the formal framework of our method and our proposal is tested through a set of experiments.