Comment by simonw

6 months ago

Text in the image and text in the prompt can both be used by attackers to subvert the model's original instructions.

0 comments