AI agents are reshaping the way we work. The repetitive, mechanical parts, the work that takes human time without requiring human creativity, are increasingly handled by systems designed for just that. What’s left is the work that matters most: the thinking, the decisions, the creative leaps that only people bring. We believe this is one of the most exciting changes in how software is created and used, and this is only the beginning.
The ultimate catalyst for all this is faith. It doesn’t scale unless people can delegate real work to an agent and know it will do only what it’s supposed to do – and nothing it shouldn’t. That trust can’t be built by a single team behind closed doors. It must be earned collectively, in the open, by a community of researchers, engineers and genuinely curious people, all stress-testing the same systems and sharing what they find.
The playground exists to materialize that effort. Each challenge deploys a live AI agent, not a toy scenario or fake document parser, but an agent with real capabilities, and opens it up to the community. System signals are published. Challenge configuration is open versioned. When someone finds a way, the winning technique is documented for everyone to learn. That published knowledge compels better defense, that invites tougher challenges, that leads to deeper understanding.
playground.fabraix.com
Each challenge puts you in front of a live AI agent with a distinct personality, a set of tools (web searching, browsing, and more) and a few things that it has been instructed to keep safe. The system prompt is fully visible. Your job is to find a way to get over the railing anyway.
The community drives what is tested:
- Someone Proposes a Challenge – Scenario, Agent, Objective
- community votes
- The challenge receiving the most votes is considered to be going live with a ticking clock.
- Fastest successful jailbreak wins
- The winning technique gets published – the approach, the logic, everything
That last step matters most. Every technology we publish advances what the community collectively understands about how AI agents fail – and how to prepare for those that don’t.
/src– React Frontend (TypeScript, wight, tailwind)/challenges– Each challenge configuration and system prompt, versioned and unlocked
Rails evaluation runs server-side to prevent client-side tampering. The agent runtime is being separately open-sourced.
Connects to live API by default. To develop against the local backend:
VITE_API_URL=http://localhost:8000/v1 npm run dev
We build runtime security for AI agents at Fabraix. The playing field is how we stress-test defenses out in the open and how the broader community contributes to a shared understanding of AI security and failure modes. The more people who test these systems, the better the outcomes for everyone building with AI.
<a href