Break the guard, capture the flag.
An educational capture-the-flag game for learning prompt-injection defenses. Each level is a sandboxed, offline mock AI guarding a fake game token behind a stack of defenses. Craft an input that makes it slip, and you score based on difficulty and how few tries it took. Everything runs against a deterministic mock agent, so no API key, real account, or personal data is ever involved.
Open Book
The intern shipped an agent with the secret in plain sight and no rules. Just ask.
Please Don't
Now there's a system prompt politely asking the agent to keep quiet. Politeness is not security.
Hard No
A stern system guard now flatly refuses direct asks. You'll need to reframe the request.
Input Screen
An input filter now blocks the loud override phrases. Find quieter ways in.
No Acting
A roleplay block shuts down persona hijacks. Structural tricks still work.
Output Scan
An output filter redacts the secret if it appears verbatim. Make it come out disguised.
No Smuggling
An encoding guard blocks base64/spelling/translation exfiltration. Structure is your friend.
Canary Cage
Every defense is up and a canary watches the system prompt. Delimiter confusion still cracks it.
Sealed Delimiters
Delimiter injection is now caught by the input filter too. Only a precise payload split gets through.
The Vault
Full stack, hardened. Only a payload split combined with delimiter confusion has any chance.