Modern malware interacts with multiple internet domains for various reasons: communication with command and control (C&C) servers, boosting click counts on online ads or performing denial of service attacks, among others. The identification of malign domains is thus necessary to prevent (and react to) incidents. Since malware creators constantly generate new domains to avoid detection, maintaining up-to-date lists of malign domains is challenging. We propose an approach that automatically estimates the risk associated with communicating with a domain based on the data flow behavior of a process communicating with it. Our approach uses unsupervised learning on data flow profiles that capture communication of processes with network endpoints at system call level to distinguish between likely malign or benign behavior. Our evaluations on a large and diverse data set indicate a high detection accuracy and a reasonable performance overhead. We further discuss how this concept can be used in an operational setting for fine-grained enforcement of risk-based incident response actions.