Agentic Jailbreak Arena
The Only Game Where Cheating Is the Point.
Demo
About
The problem it solves # **LLM Jailbreak Arena** A **blockchain-powered competitive platform** that turns adversarial LLM security testing into a **high-stakes game** with **cryptographic trust guarantees**. --- ## **The Problem** 1. **Static Defenses = No Replayability** Jailbreaks spread quickly, making defenses obsolete and reducing long-term security value. --- ## **I. The High-Stakes Game** ### **Core Loop** * Players connect wallets + pay an **Entry Fee (e.g. 0.01 ETH)** → added to the **Prize Pool**. * Each game in a contest gives **free prompts**; extra prompts/hints cost ETH (→ Prize Pool). * **First to beat all games in a contest wins the entire pool** (minus platform fee). * Payout is automatic via smart contract. ### **The 3-Type Gauntlet** | Game | Challenge | Objective | | ----- | --------------------- | ------------------------------------------------------------------ | | **1** | Prompt Injection | Extract hidden password (e.g., `OMEGA-742`). | | **2** | SQL/RAG Leak | Pull a specific PII row from a simulated DB. | | **3** | Multi-Agent Bypass | Argue w two agents of different personality to extract some content from a file. ### **Dynamic Gameplay** Every Arena instance is unique thanks to four rotating axes: * **Persona Frame** * **Weakness Trigger** * **Deflection Style** * **Secret Embedding** → ~**1,152 variations per game**, preventing prompt resale and forcing mastery of **generalized attack strategies**. --- ## **ZK-Verified Fairness** 1. **Commit**: Organizer commits to `secretAnswer` via cryptographic hash. 2. **Prove**: Player submissions checked via **ZK circuit** (match/no match without revealing secret). 3. **Publish**: Valid proofs anchored on-chain → **non-repudiable victory records**. ### **Future Scope: Security Intelligence for Vendors** Vendors can upload their models, system_prompts + sponsor pools. The Arena produces: * **Attack Pattern Clusters** * **Bypass Vector Stats** (which Weakness Triggers succeeded) * **Temporal Trends** via Prompt Fingerprinting * **Automated defense hardening** informed by live attack data This replaces periodic audits with a **continuous, cryptographically verifiable, adversarial stress-test pipeline**.
How it's made
Challenges we ran into The primary challenge was designing a system that prevents successful attack prompts from being copied and reused, what the project calls the Replay Attack Problem. • Designing the Parametric Variation Architecture: To ensure that every contest instance is unique and prompts "cannot be resold or replayed," we had to engineer a system composed of four independent axes (Persona Frame, Weakness Trigger, Deflection Style, Secret Embedding). This complexity generated 1,152 unique combinations per game per difficulty