The paper presents OutGene, an approach for streaming detection of malicious activity without previous knowledge about attacks or training data. OutGene uses clustering to aggregate hosts with similar behavior. To assist human analysts on pinpointing malicious clusters, we introduce the notion of genetic zoom, that consists in using a genetic algorithm to identify the features that are more relevant to characterize a cluster. Adversaries are often able to circumvent attack detection based on machine learning by executing attacks at a low pace, below the thresholds used. To detect such stealth attacks, we introduce the notion of time stretching. The idea is to analyze the stream of events in different time-windows, so that we can identify attacks independently of the pace they are performed. We evaluated OutGene experimentally with a recent publicly available dataset and with a dataset obtained at a large military infrastructure. Both genetic zoom and time stretching have been found to be useful, and high values of recall and accuracy were obtained.