Loops HouseStand-in

NEAR HACKATHON
Links
Team
1 member- BOOwner
Bonney Mantra
Overview
AI coding agents (Claude Code, Codex) read ~/.near-credentials, .env files and shell history as part of normal work, so a prompt injection or a compromised npm package can steal a NEAR full-access key. With Stand-in the owner's key stays in their own wallet (NEAR Connect; tested with Meteor), which approves a function-call key for the agent scoped to your contract (without a wallet, the CLI moves the real key out of reach), and Stand-in plants a unique decoy key per agent session everywhere a key would live. Every decoy belongs to one bait account. When any decoy signs a transaction, a watcher freezes the real account through a guardian contract (median 3 blocks): it deletes the agent's keys, removes its keys from intents.near (proven on mainnet) and refuses its Chain Signatures, so the agent cannot sign on other chains either. Claude Code and Codex hooks name the session and tool call that leaked the decoy. An OutLayer TEE verdict puts the thief on a blocklist that agents check before paying, and a NEAR AI Cloud TEE model writes a signed, attested incident report. Benchmark: 20 attack routes, 11 of 11 thefts frozen; with Stand-in off the same attacks drained 2.4 NEAR.