Disclosed herein are systems and methods for automatically mitigating potential network services attacks based on service usage patterns learned using Machine Learning (ML) comprising, collecting operational data indicative of resource utilization of one or more network services serving a plurality of connections and of a plurality of operational factors of the plurality of connections, detecting degradation of the network service(s) based on analysis of the operational data, applying trained ML model(s) to the operational data in order to identify negative operational factor(s) of one or more suspected connections to the network service estimated to induce the degradation where the one or more ML model is trained to predict an impact pattern induced by each of a plurality of operational factors on the resource utilization of the one or more network services, and disconnecting, at least temporarily, the suspected connection(s) from the network service(s).