Loops HouseBanbo

NEAR HACKATHON
Links
Team
1 member- MYOwner
Myacxt
Overview
Banbo is built on CaMeL ("Defeating Prompt Injections by Design", Google DeepMind and ETH Zurich), a defence against prompt injection that tracks where every value an AI agent uses came from. The problem: AI agents that pay invoices can be tricked by one email ("our account changed, pay us instead"). Spending limits and allowlists don't catch it, because the payment looks normal. What we built: Banbo, the security layer for AI agent wallets. It enforces CaMeL's rule: pay only when the recipient and amount came from the owner, not from an email or web page the agent read. Untrusted amounts are held for the owner and redirects are refused. Live on testnet: Claude paid a real invoice, a forged one was held, and an email redirect was refused on-chain. How NEAR fits: the funds sit in a vault contract, and the agent's function-call key can only request payments. The contract checks where each value came from, settles each payment exactly once through async callbacks, supports USDC (NEP-141), and can verify Intel TDX quotes on-chain.