Members-Only
Recent Talks & Demos are for members only
You must be an AI Tinkerers active member to view these talks and demos.
Jailbreak RUN: Gamifying LLM Security
Participants collaboratively attempt to jailbreak an LLM‑powered bot using prompt engineering and RAG manipulation, revealing security flaws and learning LLM logic in real time.
Jailbreak RUN is both a community event and a Discord LLM-powered bot with a simple concept: challenge an entire server of users to work together to break the bot and extract a password or force it to perform a specific action. The bot is designed to hybridize prompt engineering with application logic tailored to a specific event, and through attempts to break the bot, it also enables storytelling tied to uncovering information about the game directly from the bot.
So far, the following editions have been conducted:
- A classic Jailbreak with guardrails.
- A Jailbreak with encrypted input and output messages, making some jailbreak methods more difficult or outright impossible to use.
- Jailbreaking a vector database by attempting to override the RAG (Retrieval-Augmented Generation) system to extract hidden sensitive data from the database while simultaneously bypassing guardrails.
This concept is highly scalable to other variants, such as the planned PvP mode, where players will design bots for one another and compete in a tournament to see who can jailbreak the other teams’ bots first.