Cyber attacks cause major disruptions of online operations, and might lead to data and revenue loss. Thus, appropriately training security analysts, human decision makers who are in charge of protecting the infrastructure of a corporate network from cyber attacks, on different frequencies of cyber threats (base-rates) is indispensable to improving their on-job performance. However, little is currently known about how training analysts on different cyber attacks, that differ in the base-rate of cyber-threats, affects their on-job performance in a highly dynamic environment, while confronting novel transfer conditions. We report a laboratory experiment where human participants are trained on two different cyber-threat base-rates, high and low, and are transferred to an intermediate base-rate level of threats. The experiment helps us to develop an understanding of the situational attributes that participants attend to during their detection of cyber-threats. A linear model that is based upon participants’ attended attributes and calibrated to the two base-rates during training does well to capture the performance during transfer. We use the calibrated model to generate predictions in novel real-world transfer conditions that contain a low cyber-threat base-rate and a shorter training period.