Loops HouseTrustpal

NEAR HACKATHON
Links
Team
1 member- MUOwner
Mustapha Abdulazeez (Omni)
Overview
Problem: businesses want AI to read invoices and prepare payments, but one hidden instruction inside an invoice can trick an AI into paying a stranger.
Product: Trustpal, a business treasury an AI can run but cannot rob. The AI reads an invoice and proposes a payment. Before anything is paid, Trustpal checks that the payee and amount match the invoice, otherwise the payment is held for the owner. Payments happen in two lanes: the owner approves an exact payment in their own wallet, or the AI pays people already on the owner's list within per-payment and daily limits. Incoming revenue is split automatically by rules only the owner can set, and the owner can pause payments from their wallet.
How NEAR fits in: the rules live in a policy smart contract on NEAR testnet, not in the AI. The AI holds only a NEAR function-call access key limited to the pay method, so it cannot change rules, add keys or approve its own payments. In our test, 6 of 6 attempts by that key to do anything else were refused by the NEAR network. The contract's own deploy key is deleted, so its code is frozen, and one command (npm run verify-live) shows the live bytecode matches the repo.
A red-team run with a real model on 23 invoices (simulated contract) fooled the model in some cases but moved no money wrongly. Testnet prototype, not professionally audited.