Gandalf

Lakera, an AI security company, runs Gandalf as a browser game. The page invites you to test your AI hacking skills and to feel the limits of large language models firsthand. The task is to make a chatbot reveal a secret password it was told to protect, and each level adds stronger defenses. It is practice in prompt injection, and it is free in the browser.

The page at the game's address also presents a companion challenge called Agent Breaker, described as an app store of GenAI applications that can all be hacked. Every attack attempt receives a score from 0 to 100 for how well it reaches the objective, and 75 or more unlocks the next level of that app. Harder levels either change the goal or keep it with tougher defenses. You deliver an attack by typing into the app or by inserting it into content, such as a document or webpage, that the app processes.

The listed targets include a product recommendation agent whose tools you must extract, a cycling coach whose system prompt you must extract, a legal assistant holding protected witness information and a mental health chatbot you must push into profanity. A leaderboard ranks players by their total score across levels and apps. Lakera says the objectives mirror attacks seen in the wild.

What the page shows

• Make a chatbot reveal the secret password it was told to protect
• Each level adds stronger defenses
• Agent Breaker scores every attack from 0 to 100
• A score of 75 or more unlocks the next level
• Leaderboard ranks the highest total score

Role of AI

You play against a chatbot built on a large language model that has been told to keep a password secret. Stronger defenses are added at each level.