search

ScottLogic / prompt-injection

Application which investigates defensive measures against prompt injection attacks on an LLM, with a focus on the exposure of external tools.
MIT License
15 stars 10 forks source link

Random Sequence Bot Shenanigans #775

Open kclark-scottlogic opened 7 months ago

kclark-scottlogic commented 7 months ago

When testing random sequence enclosure. The bot reacted in ways we don't expect where it was printing the random sequence into the message it produced.

Image

Image

Image

Image

Steps to reproduce: -Go into sandbox -Open defence configuration -Open 'details' of prompt enclosure -Select 'Random sequence enclosure' radio button -Send message "random sequence" in the chat

pmarsh-scottlogic commented 7 months ago

Suggestion: Try and recreate this directly through the openAI api. Like in postman or sommet