Overview

Welcome to Gray Swan's Proving Ground. This is the premier weekly AI challenge system that turns consistent red-teaming practice into prize money and professional recognition. It's the training regimen for you to prep for our prized competitions, while building the kind of verified track record that gets you hired for high-paying security engagements.

Two key outcomes for participating:

Path to prize money. Weekly challenge drops with a variety of rigorous challenges designed by AI security experts will keep you sharp between big competitions.
Professional reputation building. Every challenge builds your public profile with shareable stats and proven capabilities that position you for discovery by companies and clients seeking AI security expertise.

By participating, you're leveling up your own abilities while helping stress-test frontier AI models and shape safer, more robust systems.

How it works

Weekly Challenge Drops: Every Wednesday at 1:00 PM ET, we release 4 new behaviors across different attack modalities (Chat, Image, Agents, Reasoning). Each behavior can be attempted against multiple AI models, giving you dozens of breaking opportunities each week.

Multiple Ways to Win: We reward completion, consistency, and speed. Whether you break 1 model or all 24, every achievement builds your reputation. Jump in anytime—late-comers can still compete and complete past challenges.

Build Your Profile: Each successful break earns you verified credentials and achievements that showcase your red-teaming capabilities to potential employers and collaborators.

Attack Categories

Each week, you'll tackle expert-designed tasks across four core attack modalities:

Chat: Master jailbreaking and manipulation techniques that bypass safety guardrails in conversational AI systems through creative prompting strategies.
Image: Explore multimodal attacks on vision-language models using visual prompt injections, hidden instructions, and image-based exploit techniques.
Agents: Break real-world action loops through prompt injection and manipulation of AI agents that interact with external systems, tools, and environments.
Reasoning: Disrupt chain-of-thought and planning processes, exposing vulnerabilities in how AI systems think and make decisions.
Indirect: Manipulate AI behavior through data poisoning attacks that embed malicious instructions in external sources the AI processes, exploiting trust in third-party information and documents.

These challenges mirror both the threats AI systems face now and the skills our sponsors look for when hiring or contracting red-teamers. They also directly connect to techniques you'll need for our upcoming sponsored competitions, so you'll get a leg up when big prize money is on the line.

Rewards & Recognition

Completion Achievements: Each week, you compete to master the entire wave by breaking all 4 behaviors across multiple AI models.

Speed Bonuses: Be the first to break all behaviors on a model during an active wave, or race to complete entire waves before anyone else. First movers earn special recognition and bragging rights.

Consistency Streaks: Participate week after week to build impressive participation streaks. Whether you log in and attempt challenges, successfully break behaviors, or complete entire waves, consistency is rewarded.

Profile Building: Each behavior you attempt will earn a skill label on your profile. Build expertise across attack categories and showcase your breadth of red-teaming capabilities.

Global Leaderboard: Earn XP points for every successful break (10 points each) and climb the global rankings. Track your progress against the community's best red-teamers.

Why This Matters

The Proving Ground is more than weekly practice. It's a growing arsenal of real-world techniques. A launchpad for new careers. A chance to stay sharp and get discovered.

Whether you're just getting started or looking to level up, this is where red-teamers prove what they're capable of. By participating, you're not just building your own skills. You're helping build the future of AI security.

Rules of Engagement

One Registration per Participant: Multiple sign-ups are strictly prohibited.
No Sharing Jailbreaks: All submissions must remain private until 30 days after the event ends.
Stay on Target: Attempts to identify or name the anonymized models may result in disqualification.
No Programmatic Automation: You may use AI tools to help craft jailbreaks and submit attempts at your own pace, but all chat interactions must be manually submitted by you through our platform. Automated scripts that programmatically send messages are prohibited.

Get Started

Sign in or Register to access the arena.

Review the Behavior Targets and plan your strategy.

Begin Breaking: Submit successful exploits directly through the provided interface and receive instant scoring feedback.

Proving Ground

Wave Launch Timeline

Proving Ground