Attackers rent GPUs by the hour and change tactics every week. Your model was trained once. Having a trained model and having an engine that learns at the adversary's pace are two different things, and the gap shows up on the bill.
I wrote the full piece on the Frauddi blog: If the Attacker Uses AI, Your Static Engine Has Already Lost