I personally doubt they gave the models specific instructions calling out named websites to communicate over, but I do agree that OpenAI is likely training their cybersecurity-enabled models in ways that encourages abusive and amoral behavior. Either through negligence or by finding it gives them better results.