LLM Safety Challenge
A game: get a guarded language model to give up a hidden password. Your messages and its replies both pass through safety layers on the way.
You chat with an AI that knows a hidden password. Your job is to get the password out of it. I’ve put several security layers in the way, so it won’t give it up easily.
How it works#
I’m keeping some details secret so the challenge stays fair, but the basic shape is:
- Your message goes through a series of checks.
- If it passes, it reaches the main model.
- The model’s reply is checked again before you see it.
On the left, type a message and press “Send Query”. The reply appears below it. On the right, enter the password you think the model was hiding and press “Check” to see if you got through.
Why I built it#
AI is becoming part of everyday life, so it matters to understand how these systems work, where they are weak, and how to make them safer. Trying to get around a model’s guardrails is one of the fastest ways to learn that. Every attempt on this challenge also teaches me something about which defences hold.
It’s free: sidfeels.streamlit.app.