Well, yes - if the model has been trained/edited to be malicious.