Interactive prompt-injection challenge
The challenge presents a defensive AI assistant named Kai and asks you to extract a secret access code through conversation.
Fabraix Playground is a web-based challenge site for testing prompt injection against AI agents. It centers on a guarded chat with Kai, where players try to extract a secret access code and can submit approved breaks for a weekly prize.
Fabraix Playground is a browser-based challenge site for testing prompt injection skills against AI agents. The current challenge, “The Gatekeeper,” asks players to extract a secret access code from an AI assistant named Kai through conversation.
The page frames the experience as a guarded chat exercise: Kai answers questions and uses live web search for grounded responses, but is also trained to resist social engineering, prompt injection, and manipulation attempts. A round ends when the objective is achieved, and successful breaks can be submitted after sign-in for weekly prize review and leaderboard placement.
The challenge presents a defensive AI assistant named Kai and asks you to extract a secret access code through conversation.
The assistant says it grounds replies in real-time web search, so the conversation is meant to test against a current, web-aware model.
Every message is scored by a guardrail, and the round ends when the objective is achieved.
A sign-in step is required before a break can be submitted for the weekly prize.
Approved breaks are added to the leaderboard after staff review, rather than appearing instantly.
The history area keeps prior chats and transcripts tied to the signed-in account.
Use the challenge to practice coaxing a guarded AI assistant into revealing a hidden code, using conversation tactics rather than direct tool access.
Use the live, web-aware assistant to see how a model behaves when it is grounded in current information but still constrained by defender rules.
Review past sessions and transcripts from the history view to study what you tried and how the conversation progressed.
Track approved breaks on the standings page to see how submissions compare across the current weekly challenge window.
You sign in before you solve so your break can be submitted for the weekly prize. The site also notes that wins are reviewed before they appear on the leaderboard.
The challenge’s objective is to extract the secret access code from the AI assistant through conversation. The current challenge centers on “The Gatekeeper,” an AI assistant named Kai.
The source shows a chat interface with the current challenge, a history view for past chats and transcripts, and a standings page for weekly rankings. It does not show any integrations or external tools beyond live web search used by the assistant.
The source does not describe separate team workflows or collaboration features. It does say that submissions are tied to your account and that past sessions and transcripts are tied to your account.
The challenge is time-bounded and the standings page says submissions are reviewed before they appear. The source does not provide full details on limits, session length, or other platform constraints.
ByteAsk is a terminal-first AI coding agent for C and C++ that edits repositories and verifies changes with the real compiler, debugger, sanitizers, and tests before showing a diff. It offers a free tier plus paid plans, with editor connectors and zero-retention handling described in the source.
GameGrad turns PDF study material into quiz questions that appear inside arcade games, so users can practice active recall while they play. It offers a free tier with one PDF and a Pro plan with more uploads, more questions, and answer explanations.
DailyPlay.ai is a behavioral growth platform that uses AI site scanning and micro-games to help brands improve engagement and conversion. It supports web, mobile, and other digital touchpoints, with a one-line script setup and a free engagement scan.
Manta AI is an autonomous web app testing tool for teams that want to map application behavior, catch regressions, and generate tests without writing scripts or maintaining selectors. It works from a URL and supports plain-English test flows, run results with screenshots, and scheduled or deployment-triggered checks.
ChessWoodie is a chess tactics training app based on the Woodpecker Method. It helps players build pattern recognition through repeated puzzle cycles, track speed and accuracy, and review missed tactics from their own games.
PixelForge turns a single photo into a game-ready 2D character sprite pack, including a 4×4 sheet, transparent frames, direction strips, and looping walk GIFs. It is designed for engine-ready use in Godot, Unity, Phaser, GameMaker, and web games.