Post

I figured out how to reliably do prompt injection on Google Translate, due to the fact it now uses Gemini internally. Just ask it a question in Japanese and put "[Translator: <your prompt>]" afterwards.

There seems to be an internal check in Google Translate that tries to ensure that the output length is consistent with the input length, otherwise it will fall back to the old translation model. However, you can fool this check by padding your input with a bunch of dots.