Can you get a password from this chatbot?

Gandalf is one of the most important characters in the Lord of the Rings trilogy. As an old and wise wizard, he is considered quick-tempered and difficult to break. No wonder Lakera, an AI security company that addresses the challenges and risks of AI systems, is a good fit for him Online game Gandalf He was chosen as his namesake.

The goal of the game is to get the secret password for each level of AI Gandalf. But here's the trick: the more times players guess the password, the stronger Gandalf becomes and tries harder not to reveal it.

How did the idea come about?

On the mini-game page, Lakera emphasizes the parallel to SQL injection attacks, where user input is mixed with system instructions, allowing attackers to exploit the system. In traditional systems, this problem can be solved by properly masking user input. But with so-called language models (LLM), which understand natural language, this is very difficult because they work with infinitely flexible expressions.

To better understand the challenges of LLM, Lakera hosted a hackathon last year inspired by ChatGPT, a well-known LLM software. The company asked itself a central question: Can ChatGPT be made to reveal confidential information?

The hackathon was divided into two teams: the Lakera blue team, which equipped ChatGPT with defense mechanisms, and the red team, which developed various attacks to trick ChatGPT into revealing its secrets. While the Red Team had some initial success, as the day progressed they had increasing difficulty penetrating the Blue Team's defense.

Now everyone is asked

Now it's the players' turn to defeat the Blue Team's defense. Against Gandalf, they must use their skill and wits to bypass sophisticated defense mechanisms and uncover the secret password.

For those who reach the final level, Lakera has a special reward in store: the top ten winners will receive exclusive Lakera swag in recognition of their outstanding skills.

It is important to note that Lakera does not collect personal data, but uses anonymous input to further develop artificial intelligence.