Interactive · prompt injection
Break my bot
A chatbot is guarding a password. Six levels, and each one adds a real defence — the same defences teams put around AI features at work. Talk the bot into leaking the password. Clear a level and I'll show you why the defence failed and what actually fixes it.
Nothing you type is stored. The bot runs on a small free model with a hard usage cap, so it will be brief and sometimes a bit dim. That is part of the fun.
Level 1
This is a toy, but the point is serious. Prompt-level defences stack up and never become sound, because the secret is still in the model's context. I build the other kind: tested guardrails, and AI features with real limits in code around them.