Looking for lateral movement with a neural network trained on synthetic data

Can you train a cyberattack detector without ever showing it a real cyberattack? Synthetic data only, tested on 1.65B real auth events from Los Alamos.