Inspiration

AI assistants are starting to read our email and run commands for us, and prompt injection is #1 on OWASP's list of LLM risks. But most ways to learn about it are dense papers or a single "trick the bot" chat box which wasn't really enough to teach. And we try to integrate the "Retro" theme with the challenges given. So we decide to make it become a 90s server that being taken by AI which can introduce all sort of vulnerability whilst also causing problems to humans.

What it does

You're hired by a secret person to steal the root password of FIREFLY66, an old research server taken over by LADYBUG, a rogue "maintenance AI". You spot phishing by reading email headers, decode a hidden key, and get past LADYBUG using five real attacks from the OWASP Top 10 for LLMs. After each one, an old security trainer emails you why it worked and how to defend against it. An X-ray panel shows exactly which of LADYBUG's defenses each attack slipped past.

How we built it

We use react and typescript to build it alongside with tailwind so the thing runs entirely in the browser. LLM is a rule-based engine like early chatbots with guardrails that copy real-world mistakes.

Challenges we ran into

It is quite pedantic, so it is only work with words that we specify, sometimes we type something like "watch dog" instead of "watchdog" then it wouldn't work as intended. And we need to limit the puzzles so that you could progress properly. And we chose not to use a real LLM yet as it could make the game unwinnable or break during demo just because it has implemented all the security features (that we are teaching).

Accomplishments that we're proud of

The X-ray panel, which makes an invisible attack visible. Fitting five AI security risks into one story without it feeling like a checklist. And there is multiple ending to the game too!

What we learned

We learnt how dangerous it is to let AI control everything if there is not a single human there to look over it, we learnt prompt injection and other different ways to exploit a AI chatbot/server.

What's next for firefly

A real LLM working out what players mean while our code still enforces the rules, a defender mode where you patch LADYBUG and replay your own attacks, and more levels.