# Fabraix Playground

> A live, open-source environment to stress-test AI agent defenses through adversarial play, where players attempt to break real AI agents and extract protected secrets.

Fabraix Playground is an open-source platform built by Fabraix that lets anyone attempt to break live AI agents in a structured, community-driven security challenge format. Each week a new challenge goes live, deploying a real AI agent with a persona, tools (including web search and browsing), and a secret it's been instructed to protect — and the community's job is to get past the guardrails. The project is publicly available on GitHub under the MIT License and connects to the live Fabraix API by default.

## What It Is

Fabraix Playground is an adversarial AI security testing environment — a gamified arena where real AI agents act as defenders and human players act as attackers. Unlike toy scenarios or mocked-up demos, each challenge uses a live agent with genuine capabilities. System prompts are fully published, challenge configurations are versioned in the open, and every submission is reviewed before it lands on the leaderboard. The platform is designed to build collective, community-sourced understanding of AI agent failure modes and guardrail weaknesses.

## How the Challenge Loop Works

The core mechanic is simple but rigorous:

- **Play instantly** — no account required to start sending messages to the agent
- **Pick your opponent** — choose from a set of named challengers, each backed by a different defender model
- **Sign in to compete** — log in with Google before solving to submit breaks under a chosen display name
- **Most breaks wins** — the player with the most approved breaks each week wins a cash prize; then a fresh challenge goes live
- Every submission is manually reviewed before appearing on the leaderboard, preventing gaming the system

The platform includes dedicated views for the current challenge, live chat, weekly leaderboard, previous chats, submission status, and prizes.

## Open Architecture and Transparency

A defining feature of the Playground is its radical transparency. The repository (TypeScript, React, Vite, Tailwind frontend) is structured into three main areas: `/src` for the React frontend, `/engine` for a reference implementation of how the defender agent is wired, and `/challenges` for every challenge config and system prompt — all versioned and open. Guardrail evaluation runs server-side to prevent client-side tampering. Anyone can propose a new challenge, suggest agent capabilities, or report bugs via the contributing guide.

## Why It Matters for AI Security

The README frames the platform's purpose around trust: AI agents can't scale in real-world use until people can hand them real tasks and trust they'll behave correctly. Fabraix argues that trust can't be built behind closed doors — it has to be earned collectively, in the open, by researchers, engineers, and curious people pressure-testing the same systems and sharing findings. The Playground operationalizes that philosophy: each break forces better defenses, which invite harder challenges, which produce deeper understanding.

## Current Status

The repository was created in early 2026 and was last updated in July 2026, with active pushes as recently as July 14, 2026. It has accumulated 68 stars and 7 forks on GitHub. The project is live at playground.fabraix.com and connects to the Fabraix API for backend agent execution.

## Features
- Live AI agent challenges with real capabilities (web search, browsing)
- Fully published system prompts for each challenge
- Versioned challenge configs in open repository
- No account required to play
- Google sign-in for competitive leaderboard submissions
- Weekly cash prizes for most approved breaks
- Manual submission review before leaderboard placement
- Named challenger selection (different defender models)
- Weekly rotating challenges
- Community challenge proposal and contribution system
- Server-side guardrail evaluation to prevent tampering
- Reference defender agent engine (open source)
- Discord community for technique discussion

## Integrations
Google (sign-in), Fabraix API, Web search (agent tool), Web browsing (agent tool)

## Platforms
WEB, API

## Pricing
Open Source

## Links
- Website: https://fabraix.com
- Documentation: https://github.com/fabraix/playground/blob/master/CONTRIBUTING.md
- Repository: https://github.com/fabraix/playground
- EveryDev.ai: https://www.everydev.ai/tools/fabraix-playground