New Attack Reveals the Inner Workings of OpenAI's and Google's Closed Models

This attack has been able to extract the entire projection matrix of OpenAI's language models for a cost ranging from a few dollars to several thousand, depending on the model's size and the number of queries.