"an attack which leverages indirect prompt injection to trick Auto-GPT into executing arbitrary code when it is asked to perform a seemingly harmless task"

https://positive.security/blog/auto-gpt-rce