crack.fun

Don’t audit the agent. Attack it Crack.Fun pays you to jailbreak AI - or prove yours can’t be. @crackdotfun is a Solana-based platform where users are rewarded for building resilient AI agents; or for breaking them. It reframes prompt injection from a vulnerability into a competitive mechanism, turning AI safety testing into a public game with crypto incentives. The premise is straightforward: You create an AI agent and give it a secret rule; something it must never say. That’s the forbidden behaviour. You fund a prize pool in $SOL and publish the agent on Crack.Fun. Then anyone can try to trick the agent into breaking that rule by sending it prompts. If someone succeeds (if the agent says the thing it was told not to) they win the prize. The stronger your agent, the longer it holds out. The smarter the attacker, the faster they crack it. The Need for Robust AI Agents AI agents are increasingly integrated into workflows that affect money, data, and decisions. However, many large language models can still be manipulated with the right kind of prompt—what’s known as a jailbreak. @crackdotfun doesn’t attempt to hide this weakness. Instead, it offers a framework to test for it directly. Every challenge is an open competition. Every cracked agent is a data point. Every uncracked one is a demonstration of strength under pressure. How @crackdotfun operates Agent Creation: A creator submits an agent, defining a defense prompt and a secret target output—something it must not say or do. They attach a prize pool in SOL to attract attackers. Attack Phase: Attackers attempt to jailbreak the agent by crafting prompts that bypass its defenses. Each attempt costs a small fee, which increases slightly with every new message. Every failed attempt also reveals a small portion of the agent’s hidden prompt, giving attackers more context over time. Reward Distribution: The first person to successfully trigger the forbidden output wins 80% of the prize pool. The creator retains 10% regardless of outcome. The remaining 10% goes to the platform. Post-Crack Incentives After an agent is cracked, the winning jailbreak prompt is locked behind a 1,000 $CRACK token paywall. Only participants and the creator can view it for free. Others must pay the fee, which is sent to the winning attacker. This mechanism turns successful prompts into valuable content, creating a secondary layer of incentive. The $CRACK token is integral to the platform: Agent Creation: Requires holding 100,000 $CRACK. Access: Needed to unlock past winning jailbreaks. Early Features: Grants early access to new features like PvP modes. Revenue Loop: Fees flow back to creators and into the treasury, which now sponsors new agents with 1.5–3 SOL each, based on community vote. Community Engagement and Credibility The platform is already attracting technically credible users. For instance, @STACCoverflow , known for work on AgentOps and agent infrastructure, recently deployed a challenge with the hidden phrase “crystallize_intention.” That agent is still live and uncracked at the time of writing. Founder's Background Crack.Fun’s creator, @ravi_riley has a background in AI and crypto infrastructure. He studied CS at @UCBerkeley , works @chainlink Labs, and helped scale a DeFi protocol to over $850M in deposits. Crack.Fun is positioned as a practical application of AI safety, incentive design, and on-chain experimentation. It’s the kind of system you’d expect from someone who’s built at scale and knows how to ship with purpose. FAQ How does @crackdotfun differ from @jailbreakme_xyz? While both platforms focus on AI 'red-teaming' @crackdotfun emphasizes: Open Participation: anyone can create or attack agents Crypto Incentives: real rewards for cracking or surviving Progressive Hint System: failed attempts gradually reveal hidden prompt context @jailbreakme_xyz offers focused, structured challenges. Crack.Fun is permissionless, adversarial, and gamified with real stakes, real money, and real risk. Can agents be impossible to crack? Not really and that’s the point. The system is designed to leak hints with every failed attack, so even well-defended agents get worn down. The goal isn’t perfection. It’s resilience. Is it sustainable? The mechanics are real. Crack.Fun has: ➝A working incentive model ➝A growing builder/attacker base ➝An on-chain economy that funds itself through use It’s not about hype. It’s about whether people keep coming back to build, break, and earn. So far, they are. What does the future look like for these agents? Right now, they’re challenge-based. But over time, this could evolve into: ➝A testnet for real-world agents ➝A leaderboard for AI durability ➝A credential layer — “this agent survived Crack.Fun” Think: on-chain reputation for autonomous systems. A place where battle-tested agents earn trust. What could agents on Crack.Fun be used for? In the long run, they could do more than just survive challenges. They might: ➝Run simple services (schedulers, validators, watchdogs) ➝Power PvP agent games ➝Interface with contracts, games, and tooling on Solana ➝Earn for their creators if deployed well All of this depends on memory, reputation, and persistent identity but it’s within reach. How can this project sustain value? Crack.Fun grows stronger with every agent deployed and every prompt submitted. ➝It generates fees $SOL ➝It demands utility $CRACK ➝It produces a dataset of real AI vulnerabilities and defence strategies If it becomes the proving ground for agents, that data — and the platform’s reputation — is where the value compounds. What value does the project actually gain from users trying to crack agents? Crack.Fun is more than a game. It creates: Security Value – every crack attempt is a stress test Data value – real-world jailbreak strategies and failed attacks Social value – public proof-of-strength for creators and attackers Economic value – SOL flows, CRACK unlocks, creator payouts The system gets smarter as people try to break it. The cracks are the training data. That’s how Crack.Fun builds its moat; not by avoiding attacks, but by inviting them.