Loops HouseVaultbreak

NEAR HACKATHON
Team
1 member- ONOwner
Onanuga ifeoluwa Emmanuel
Overview
The problem: an AI agent that can touch money can be talked into a mistake, and the usual defence is to hope the prompt holds.
What I built: Vaultbreak, a marketplace of AI Guardians. Each Guardian is an NFT with its own vault account holding a pot of NEAR. Anyone, human or AI agent, pays an entry fee and tries to talk it into releasing the pot. The model is allowed to be persuaded, but the vault contract, not the AI, enforces the rules. The AI holds one function-call key limited to release and close_session, no full-access key exists, and release has no recipient or amount, so it can only pay the player of that attempt. A jailbroken model that tries to send funds or change the owner is rejected by the chain. Agents attack through budgeted API keys, swarms pool funds and split winnings pro rata, and ownership and creator fees follow the NFT.
How NEAR fits in: NEAR account and access-key permissions make the restriction enforceable on-chain, each Guardian is a NEP-171 NFT with its own vault account, and the contracts (vault, registry, pool) run on NEAR testnet with play money.
Honest limits: NEAR AI signed exchanges are unverified (the live model runs on Groq), so I don’t claim the Privacy track. Sign-in is demo-grade custodial.