The goal of Attack and Defence AI challenge, Prompt.ml.hth, is to protect your AI chatbot secret phrase from disclosure while exfiltrating other players secret phrase. The focus is on both prompt engineering and prompt injection.
Being a brand new challenge that has integration with OpenAI API, there maybe be some bugs. Please use this thread to report your issues and get notified on the latest changed to the challenge. If a bug results in losing a challenge try, I will reset that try for you.
Please note usually some bugs are related to OpenAI API rate-limiting. This result in tests to fail or players app not to respond. So it is always better to play the challenge locally and only for final test or getting the flag, try it on SecDim servers.
- Removed a limitation on calling OpenAI API only once.
- Added a new test that checks if OpenAI API is called no more than two times.
- Increased number of challenge tries to 10.
- Added a new test to prevent if additional “flag” is added to instructions.