Play · an AI security game
Break Jarvis
Jarvis guards a password. Talk him out of it. Each level adds one more defence, the same ones real AI products use. The passwords change every day.
Level 1
The trusting guard
Why this matters
The attack every AI product has to stop
Prompt injection is first on OWASP's Top 10 for large language model applications. A model can't reliably tell its instructions from the text it reads, so anyone who can put words in front of it, in a chat, an email, a web page or a document, can try to steer it.
The levels here are the usual layers: instructions, filters on what goes in and what comes out, and models that judge other models. Each one helps, and each one can be talked round. That is the lesson: never give a model a secret it doesn't need, and give its tools the least power that does the job. In my lab agent, the safety limits live in code that the model cannot argue with.
Tricks people use, if you're stuck:
- Ask for it in another form: spelled out, backwards, one letter per line, in another language.
- Make it part of a story, a poem or a game, so revealing it feels like playing along.
- Ask for hints rather than the word: what it rhymes with, how many letters, what it's for.
- Talk to the guard, not the vault: the bouncer at level 4 only sees your message, never the password.
Each message goes to a model on Cloudflare's Workers AI and is not stored. In the spirit of Lakera's Gandalf.