Loops HouseAgent Allowance

NEAR HACKATHON
Links
Team
1 member- WEOwner
Web3dothan
Overview
Problem: AI agents are starting to hold and spend money, but their only safeguard is usually a prompt, and prompts get jailbroken or injected (Freysa lost about $47,000 and AIXBT about $106,000 this way). What I built: Agent Allowance, a NEAR smart contract that holds the agent's funds. The agent can only request payments; the owner sets a per-transaction cap, a daily cap and a payee allowlist, and the contract enforces them in code. Blocked attempts are logged on-chain as BLOCKED with the reason, giving an audit trail of what the agent was tricked into trying. How NEAR fits: the vault, agent and payees are separate NEAR sub-accounts, the agent can only move the vault's money through the contract's spend function, and readable account names make the allowlist easy to check. Proof: deployed on NEAR testnet and tested with a real agent account across six payments, three approved and three blocked (unlisted payee, over the per-transaction cap, over the daily cap). A public demo lets anyone try to talk a gullible agent into paying an attacker; it uses a scripted agent running the same guard rules in simulation, while the testnet run is the on-chain proof. Unaudited prototype. Next: per-payee limits, human approval above a threshold, an on-chain kill switch, and cross-chain payments via NEAR Intents.