We show that giving an LLM an interface to other applications, like search, can have critical security implications. When prompt injection is used and delivered

We show that giving an LLM an interface to other applications, like search, can have critical security implications. When prompt injection is used and delivered by adversaries instead of the user themself, bad things could happen- and as far as we know, this scenario has not been studied until now.