Just fine tune the model? Prompt leakage and bypass is only a problem with zero/few shot inference, and is trivially solved by removing the need for a prompt.
Could you fine tune a GPT-3 model to do the equivalent of the "Translate this from English to French: TEXT" example?