Loops HouseScrip

NEAR HACKATHON
Team
1 member- THOwner
Theodhore. R
Overview
AI agents now pay for APIs, compute and goods on their own, and they can be talked into paying the wrong party. In our trials of four small open models (480 decisions), one injected line in an invoice made three of four models pay the attacker. Budgets written in agent code or prompts do not hold either: a plain budget contract hit with 30 parallel 0.1 USDC payments against a 1.0 USDC budget paid out 1.4 USDC.
Scrip is a spending mandate for AI agents on NEAR. The owner deposits USDC into a mandate with an allowlist of recipients, a per-payment cap, a total budget, an expiry and a guardian approval for large amounts. The agent never holds tokens or a full-access key: it holds a NEAR function-call access key that can call only request_payment on the Scrip contract. Every rule is checked inside the contract; a refused payment returns a reason code and emits a NEP-297 event, so every refusal is a verifiable transaction. Under the same 30-payment burst, Scrip paid exactly 1.0 USDC. Eight of the fooled model decisions (two per model, four models) were replayed on testnet: a plain wallet executed all eight, Scrip refused all eight.
NEAR's role: function-call access keys make the agent's key safe to leak (we publish it at scrip-dev.vercel.app/try and let anyone try to steal from it); the contract enforces the mandate; NEP-297 events and the public indexer let anyone re-check all 39 recorded claims at /verify. Attack catalog: 1,999 cases in 12 categories (1,894 run, 105 skipped, 0 breaches). Live on NEAR testnet with a gasless relayer.