Post #1361540

@connected The only viable approach to prompt injection attacks is to ensure the LLM that can run actions never gets plain text of inputs. What I should do is schedule another LLM that is constrained to just returning a priority value (0 to 10) (with a clear history for each email) then a script (JS/shortcut etc) uses to sort each email based on its ID and priority value. Then the worst a prompt injection can do is get an email to be higher priority than it should be.