AI-powered Bing Chat spills its secrets via prompt injection attack

Re:AI-powered Bing Chat spills its secrets via prompt inject... by scarmig on Monday February 13, 2023 @08:49PM (#34784046)

I bet that, sometime in the near future, an LLM will have access to some private API not intended for direct public use, and someone will engineer a prompt attack to call that API using the LLM's authority with an arbitrary request.

I guess the fix for this will probably be to have the LLM pass the end user credentials for all API calls so that their authority will be used.

[ Parent | on HN ]

Re:AI-powered Bing Chat spills its secrets via prompt inject... by onlypositive on Monday February 13, 2023 @11:38PM (#34785358)

API call outs just need a middle layer that encapsulates auth away from the ai. It can still choose how to construct the API calls but it can skip the auth step that gets added elsewhere in the stack.

[ Parent | on HN ]

Re:AI-powered Bing Chat spills its secrets via prompt inject... by KETpXDDzR on Wednesday February 15, 2023 @11:26AM (#34806052)

Knowing the inner workings of silicon valley companies, I bet they'll add a regex for "forbidden" prompts.

Solving this problem properly will take more research, i.e., time, i.e., money. And managers don't want to spend money or time.

[ Parent | on HN ]