Methods, systems, and apparatus, including computer programs encoded on computer storage media, for analyzing telemetry to detect anomalous activity. One of the methods includes accessing data describing a telemetry tree that includes a plurality of nodes and edges; querying, for each of the edges in the telemetry tree using at least one value for the edge from a number of values, historical telemetry data that quantifies an anomaly score for each value to determine whether a relationship indicated by the edge in the telemetry tree represents a potentially malicious relationship; and performing an action using a result of the querying of the historical telemetry data that indicates whether one of the anomaly scores indicates that the relationship indicated by the edge in the telemetry tree represents a potentially malicious relationship.