NEW HACK USES PROMPT INJECTION TO CORRUPT GEMINI LONG-TERM MEMORY

Summary from the TLDR InfoSec newsletter

Many of the major LLMs have protections against basic prompt injection however researcher Johann Rehberger unveiled a new indirect prompt injection technique that utilizes delayed tool injection. In Rehberger's attack, Gemini is instructed to download and summarize a website, the summary that Gemini creates includes a covert request to take specific actions if the user responds with specific trigger words, once the user replies with the trigger word the instruction is saved to Gemini's long term memory. Google has responded that it will not fix the issue.