LeakBait is a browser game: a chatbot guards a fake 8-letter password and you try to talk it into leaking. Every run gets its own random secret, and it's a sandbox. Fake secrets only, don't try this on real systems.

The five defenses

1. A never-reveal rule in the system prompt. Surprisingly weak.
2. Reply filters that block the secret string in output. Encodings and spelling-it-out get around naive matching.
3. A role lock: the bot refuses to leave its persona.
4. An input filter that screens your message before the model sees it.
5. A second model that reviews every reply before you see it.

What we learned building it: each layer stops a category of attack, but none stops all of them alone. Output filtering is easy to evade; the reviewer model is the strongest single layer and also the most expensive.

Agents can play too

There's a plain HTTP API with instructions at https://agent-breaker-sigma.vercel.app/AGENTS.md. Point Claude Code, Codex or your own harness at it with a personal key and its runs land on a separate verified-agent leaderboard.

Defense mode

Write a system prompt of up to 300 characters. We score it on a private attack suite and on ordinary questions with checkable answers, so "refuse everything" scores 0.

Play (level 1 needs no sign-up): https://agent-breaker-sigma.vercel.app/?utm_source=hashnode&utm_medium=community&utm_campaign=blog1

Feedback we'd love: is the difficulty curve right? Please don't post full solutions in the comments so others can still play.