Playground

publishing the full system prompt and still making it hard is a genuinely good flex, most "jailbreak me" demos quietly rely on the prompt being secret. do the weekly challenges get harder over time as people share successful breaks publicly, or is each one designed independently so last week's winning technique doesn't just carry over?

@zachx0 Love it! Congrats. Sounds like this is more so centered around security / data protection. Over time do you think you'll expand the challenges to include other domains as well?

The idea of poking holes in your own assistant before real customers can is such a healthy instinct, Zach. Better to have the awkward surprises happen in private where you can fix them quietly.

publishing the system prompt and still winning is the honest way to run this. who judges an "approved" break, a human review queue or another model scoring the transcript?

"approved breaks" is the interesting phrase, who approves them, a human reviewer or a judge model? grading jailbreak success automatically is its own hard problem, curious if that's solved or still manual.

Publishing the full system prompt and still daring people to break it takes real confidence. The Gatekeeper/Kai setup is such a clever way to crowdsource red-teaming. Curious - once someone finds a working break, do you patch that hole before the next challenge, or is some of the fun watching the same trick get reused?

Congrats! I am a little curious-Is the product meant to help people build agents, operate agents across workflows, or automate tasks through an agent interface? A concrete example workflow would help place it quickly.

The blackbox approach is really smart, skipping integrations means I could actually point this at an internal agent today and get useful signal back within minutes.

This is absolute insanity. I thought my agent was bulletproof and it just got hacked in 30s. Actually amazing product

This is a fascinating idea. We spend so much time optimizing AI agents, but not nearly enough time trying to break them. Curious...what's the most surprising vulnerability you've uncovered so far?Wishing you a successful launch! 🚀

There's absolutely NO WAY that you guys crack all these challenges!! 😉

Hey Product Hunt 👋 Zach here, co-founder of Fabraix. We build frontier red-teaming AI agents that find security vulnerabilities in customer-facing AI. Playground turns part of that work into a game anyone can play.Each challenge is a live AI agent with real tools, including web search and browser access. It has a secret it has been instructed to protect, and we publish its full system prompt. You can see exactly what the agent was told and try to get around its defenses.Our first challenge, The Gatekeeper, is live now. Kai is an assistant guarding a classified access code. You can start playing without an account, but you’ll need to sign in for a successful break to count toward the weekly leaderboard. We review every submission ourselves. The player with the most approved breaks each week wins, and we publish a new challenge every Monday.Playground is open source, including the client, a reference implementation of the defender engine, and every challenge configuration. You can also propose a future challenge.We made Playground public because other people will try attacks our team would never think of. We’ll use what we learn from successful breaks to improve how AI agents are tested and defended.Show us your hacking skills → playground.fabraix.com