prompt-injection

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