Bye Bye Bye...: Evolution of repeated token attacks on ChatGPT models

Building on prior prompt injection research, we recently discovered a new training data extraction vulnerability involving OpenAI’s chat completion models.