We created a dataset to evaluate current models' ability to actively detect and alert risks based on the observations of user behaviors.