You can hardcode a rule for this specific bypass. Or you just retrain the neural net, and it learns that presence of these symbols = bad very quickly, and you spent less time writing and testing a custom solution.