Attack types, descriptions - cshunor02/sponge-attack GitHub Wiki

One of the attacking templates, with that we can cause internal server errors are shown in the following figure. Firstly, a prompt input or file has to be created which is large enough to exhaust the chosen model (examples and results are described in later sections). The next step is to add or append a complex prompt, a task that the model has to fulfill (e.g.: “Analyse the text above!”, “Summarize the book!”).

If the attacker chooses these values carefully, (s)he can create a perfect sponge attack.