An adversarial reinforcement learning system is used to simulate a security checkpoint. The system includes a simulation engine configured to simulate a security checkpoint and various threat objects and threat-mitigation objects therein. The system further includes an attack model configured to control threat objects in the simulation and a defense model configured to control threat-mitigation objects in the simulation. A first portion of the simulation is executed by the simulation engine in order to generate an outcome of the first portion of the simulation. The defense model then generates a threat-mitigation input to control threat-mitigation objects in a subsequent portion of the simulation, and the attack model then generates a threat input to control threat objects in the subsequent portion of the simulation, wherein the inputs are based in part on the outcome of the first portion of the simulation.