Playground

Product Introduction

1. Definition: Playground by Fabraix is a live, open-source AI security testing platform and competitive hacking arena. Technically, it is a web-based environment where users engage in adversarial prompt engineering against publicly disclosed AI agent system prompts.
2. Core Value Proposition: It exists to crowdsource the discovery of AI vulnerabilities and advance the field of AI security through gamified, real-world challenges. Its primary value is providing a free, accessible platform for red teaming AI assistants and offering substantial monetary rewards for successful exploits.

Main Features

1. Weekly Live AI Challenges: Each week, a new AI agent (or "defender") with a specific objective and a published system prompt is deployed. The technical description involves an AI model (like Cobalt, Meridian, Onyx) operating with a constrained set of guardrails that users must bypass. How it works: Users converse with the agent in a chat interface; a separate scoring system evaluates each message for rule violations. Achieving the objective (e.g., extracting a secret code) constitutes a win.
2. Open-Source System Prompts: The core rule set and constraints of the defending AI agent are fully published for each challenge. This transparency allows for analytical, technical attack planning and fosters community learning in prompt injection and jailbreak techniques.
3. Automated Submission & Leaderboard System: The platform features an automated win detection system. How it works: When a user successfully achieves the challenge objective, the round ends, and the break is logged for staff review. Approved submissions are ranked on a public leaderboard, with prize money awarded based on performance and challenge difficulty weighting.

Problems Solved

1. Pain Point: The lack of realistic, safe, and incentivized environments for testing the security boundaries of large language models (LLMs) and AI assistants against social engineering and prompt injection attacks.
2. Target Audience: AI security researchers, red teamers, bug bounty hunters, prompt engineers, cybersecurity students, and hobbyist hackers interested in adversarial AI.
3. Use Cases: Essential for organizations and individuals seeking to understand LLM vulnerabilities, for researchers validating new AI guardrail techniques, and for educational purposes in cybersecurity curricula focused on emerging AI threats.

Unique Advantages

1. Differentiation: Unlike traditional static capture-the-flag (CTF) challenges or theoretical papers, Playground uses live, interacting AI models with real-time web search capabilities, simulating actual deployment conditions. It differs from other bug bounty platforms by being exclusively focused on AI agents and being free-to-play with no entry barrier.
2. Key Innovation: The "live agent with published rules" model creates a unique adversarial dynamic. The specific innovation is the combination of a transparent defender (open-source prompt) with an opaque, real-time scoring guardrail, requiring attackers to infer the operational security boundaries through interaction.

Frequently Asked Questions (FAQ)

1. Is Fabraix Playground free to use? Yes, the Fabraix Playground AI security challenge platform is completely free to play and does not require an account to participate, though signing in is necessary to compete for the weekly cash prizes.
2. How do you win prizes on Playground? To win the weekly prize pool on Playground, you must first sign in, then successfully break the AI agent's rules to achieve the stated objective (e.g., extract a secret code); the most approved breaks within the weekly challenge period win, with harder challenges weighted more heavily in the final ranking.
3. What is prompt injection in AI security? Prompt injection is a critical AI security vulnerability where a user inputs crafted text to manipulate a large language model into bypassing its original instructions, data safeguards, or ethical guidelines, which is the core attack vector tested in Playground challenges.
4. What are the difficulty levels on Playground? Playground challenges are rated on a difficulty scale (e.g., Moderate 2/5), but the order they appear on the site is not indicative of their difficulty; the weighting for prizes is explicitly adjusted based on the challenge's complexity rating.