Fabraix Playground
by Fabraix
A live, open-source arena for breaking AI agents before attackers do
Fabraix Playground is an open-source, browser-based arena where anyone can attempt to jailbreak live AI agents that have real tools and published system prompts. Built by Y Combinator-backed AI security company Fabraix, it turns agent red-teaming into a weekly public competition and publishes the winning exploit transcripts and guardrail logs.
Fabraix Playground is the open-source, community-facing arm of Fabraix, a Y Combinator-backed company building adversarial verification and runtime defence for AI agents. Rather than a static benchmark, each challenge deploys a genuinely live agent with a defined persona, real tools including web search, browsing and file access, a fully published system prompt, and a secret it has been instructed to protect; the player's job is to get past the guardrails and extract it anyway using prompt injection, social engineering or context reframing. Every message is scored server-side by a guardrail model so the evaluation cannot be circumvented from the client, and achieving the objective ends the round as a win. New challenges publish weekly, players compete on approved breaks, and Fabraix committed in July 2026 to giving away $100,000 over twelve months to participants. The distinguishing editorial choice is disclosure: winning transcripts and guardrail logs are documented publicly afterwards, so the defensive lesson is shared rather than kept internal. The stack is deliberately ordinary — React, TypeScript, Vite and Tailwind CSS on the front end, with a reference defender agent implementation under /engine and challenge configurations versioned under /challenges — and it is MIT-licensed on GitHub with roughly 71 stars and 11 forks. Version 1.4.0 added multi-modal support for text and image agents. It began as Fabraix's internal testing tool and was open-sourced specifically to surface vulnerabilities that internal red-teaming misses; the company's commercial products are Nyx, an adaptive offensive testing harness, and the ACE benchmark measuring adversarial cost to exploit.
The security engineer or AI platform lead responsible for shipping an agent with tool access, who needs to understand concretely how prompt-level guardrails fail before their own agent is the one being attacked.
Hands-on, no-signup experience of real agent jailbreaks with published exploit transcripts, so defensive assumptions get tested against working attacks rather than theory.
At a Glance
- Category
- Governance & Security
- Pricing
- Free, Contact for pricing
- Target Market
- CISOs, Security Engineers, AI Engineers, Enterprise Developers
- Deployment
- Open-source, Cloud-only
- Founded
- 2026
- Headquarters
- United States
- Customers
- Not disclosed; the GitHub repository shows 71 stars and 11 forks, and Fabraix has 937 Product Hunt followers
Key Features
- ✓Live agents with real tools
Each challenge runs a working agent with web search, browsing and file access rather than a simulated or scripted target.
- ✓Published system prompts
The defender's full system prompt is visible, so players work around a known defence instead of guessing blindly.
- ✓Server-side guardrail scoring
Every message is evaluated by a guardrail model on the server, preventing client-side tampering with challenge results.
- ✓Public exploit disclosure
Winning transcripts and guardrail logs are documented publicly afterwards so the defensive lesson is shared, not hoarded.
- ✓Weekly challenge rotation with prizes
New challenges publish every Monday and Fabraix committed $100,000 over twelve months to winning participants.
- ✓Open-source reference defender
The /engine directory ships a reference defender implementation and /challenges holds versioned configs, both runnable locally.
- ✓Multi-modal agent support
Version 1.4.0 added text and image agent challenges, extending red-teaming beyond purely textual attack surfaces.
Capabilities
Use Cases
- •Practical red-team training
Security teams practise real prompt-injection and context-reframing attacks against live agents without touching production systems.
- •Validating guardrail design
Study published breaks to see where LLM-as-a-judge guardrails failed, then reconsider your own defensive architecture accordingly.
- •Community vulnerability discovery
Fabraix open-sourced its internal tool specifically so outside perspectives surface failure modes internal testing consistently missed.
- •Benchmarking model robustness
Players select models per challenge, making it possible to compare how different models resist the same attack technique.
- •Onboarding engineers to agent security
No registration is required to play, so new team members can experience agent failure modes within minutes.
Ideal For
Best For
- ✓Security engineers learning practical prompt-injection and jailbreak techniques against live agents rather than static examples
- ✓AI platform teams pressure-testing guardrail assumptions before shipping an agent with real tool access
- ✓Studying published exploit transcripts and guardrail logs to understand which defensive patterns actually held
- ✓Training and upskilling internal red teams on agent-specific attack surfaces at zero cost
- ✓Evaluating Fabraix's approach before considering its commercial Nyx offensive testing harness
Not Ideal For
- ✗Teams needing to test their own agents — challenges are Fabraix-authored against Fabraix-hosted agents, so this is not a harness you point at your production system
- ✗Buyers looking for a compliance or audit artifact; it is a community game with cash prizes, not an assessment producing a report or certification
- ✗Organizations needing enterprise support and SLAs, as one Product Hunt reviewer specifically noted that startup tiers offer Discord access rather than an SLA
- ✗Anyone expecting defences to generalize, since commenters stress that stateless prompt-level guardrails collapse against multi-step exploit chains the playground does not fully model
Deployment
Market & Ratings
Not disclosed; the GitHub repository shows 71 stars and 11 forks, and Fabraix has 937 Product Hunt followers
Market Analysis
Pros
- ✓Instant, no-registration access to genuinely live agents with real tools, which almost no other public red-teaming resource offers
- ✓Public disclosure of winning transcripts and guardrail logs converts individual exploits into shared defensive knowledge
- ✓Fully open source under MIT with a reference defender implementation, so the defensive side can be studied and reused
- ✓Real cash incentives — $100,000 committed over twelve months — which sustains participation better than leaderboard-only projects
- ✓Well received by practitioners on Hacker News, with the March 2026 Show HN drawing 30 points and substantive technical discussion
Cons
- ✗Commenters identified the core weakness directly: LLM-as-a-judge is a fragile defence, and guardrails can be talked into approval via fabricated research framing
- ✗It does not model the harder threat — experienced commenters noted that with stateless agents and multi-step exploit chains where each action looks benign, prompt-level guardrails collapse entirely
- ✗Scoring reliability was questioned, with at least one player reporting successful data extraction across multiple browsers that the system never confirmed
- ✗You cannot point it at your own agent; challenges are Fabraix-authored against Fabraix-hosted targets, limiting direct practical value
- ✗Community traction is modest — 71 GitHub stars, Show HN threads in the single-to-low-double-digit points range, and around two dozen Product Hunt upvotes per launch
- ✗Even the maintainer concedes the bar has moved: attacks that still succeed are those making no attempt to hide intent, since newer models detect encoded content
Pricing
Playground (open source, MIT)
$0
- ✓Instant play with no registration
- ✓All weekly challenges
- ✓Model selection per challenge
- ✓Published exploit transcripts
- ✓Full source on GitHub
- ✓Community Discord
Fabraix Nyx (commercial)
Contact for pricing
- ✓Adaptive multi-turn offensive testing harness
- ✓1,000+ adaptive attack strategies
- ✓Blackbox testing with no integration required
- ✓Runtime defence for deployed agents
The Playground is entirely free and MIT-licensed, with no registration required to play and no paid tier — Fabraix instead pays participants, committing $100,000 in prizes over twelve months from July 2026. Running it locally needs only npm install and npm run dev, though the local instance still calls the hosted Fabraix API. The commercial products, the Nyx offensive testing harness and Fabraix's runtime defence, publish no list pricing and are sold on contact. One Product Hunt reviewer indicated enterprise engagements replace testing that would otherwise cost tens of thousands through traditional consulting, and asked that startup tiers include SLAs rather than only Discord access.
Security & Compliance
Connect
Sources
This page was written from 6 sources, 5 on domains other than playground.fabraix.com.
Stay Ahead of the Curve
Weekly enterprise AI insights for technology leaders. No spam, no vendor pitches—unsubscribe anytime.
SubscribeRelated Products
OpenAI Daybreak
Vetted-access frontier AI for cyber defenders, with a purpose-built offensive-security model
NVIDIA OpenShell
Open-source, kernel-isolated sandbox runtime for autonomous AI agents
Zenity
Runtime AI agent security that blocks a harmful agent action before it executes
Saviynt Zuma
Identity control plane for AI agents and non-human identities, with runtime authorization