Good article. The real adversarial attack on GPT3 or GPT4,5,6 etc would be to hide information from the Internet. What is not in the Internet, will not be part of the training set for the GPT models.