We just published a paper showcasing newly enabled attacks on application-integrated LLMs (language models extended with other interfaces, for example search). Anyone building products by integrating LLMs right now should look at this...