Loops HouseBaitNet

NEAR HACKATHON
Team
1 member- DYOwner
Dylan Wu
Overview
BaitNet uses AI decoys to get payment addresses from scammers, and makes AI agents and NEAR Intents market makers refuse to pay them.
Problem: crypto fraud caused $9.3B of reported losses in 2024 (FBI IC3). Scam addresses surface only after victims pay, and wallets, intent solvers and AI agents share no pre-payment check.
How it works:
- A scammer AI (48 fictional scripts) talks to a decoy AI playing a victim; humans can take over either side or whisper.
- The decoy plays a persona card and thinks before replying; code checks it for AI tells and contradictions. The scammer has a mask, an operator card and a suspicion score: it tests, sends fake addresses or vanishes.
- Addresses (NEAR, EVM, Tron, BTC) are extracted live. Exchange, bridge and token addresses are blocked; doubtful ones (sent after "are you a bot?", network mismatch, contracts, very active addresses) get warnings a human must clear. A human approves every alert.
- Approved alerts go to an AlertRegistry contract on NEAR testnet (address, chain, level, evidence hash; no personal data).
- A guard returns allow/warn/block: the agent wallet refuses to sign or asks a human; a simulated Intents market maker refuses to quote.
NEAR specifics: the payment agent's only key is a function-call key for vault.pay. The Vault calls registry.check cross-contract and fails closed, so a prompt-injected agent paying a flagged account fails on chain.
Privacy (track 03): By default, LLMs run on NEAR AI Cloud TEE models; live attestation passes 18/18 checks. A zero-retention mode writes nothing on our server (covered by tests).
Live eval (DeepSeek V4.1 Flash, 48 scripts, suspicious scammers): 42 addresses, 0 leaks, 1 hijack, 13/13 poison filtered.
Runs with no keys: pnpm install && pnpm demo (local NEAR sandbox). Most code was written by Claude Code; the author made product decisions and tested it.
Judges' demo login: https://baitnet-production.up.railway.app/#/login?code=N43GTKFGEBSZ