Loops HouseRugGuard

NEAR HACKATHON
Links
Team
1 member- EMOwner
Emery Doe
Overview
AI trading agents that read the internet can be prompt-injected into buying rugs, and a check inside the prompt is only a suggestion. RugGuard enforces the check on NEAR. The agent account holds only a function-call access key scoped to one contract and two methods (request_check, execute_buy): it cannot transfer NEAR, add keys or change limits. The money sits in the RugGuard contract: a fee budget (pays for checks) and a trade budget (leaves only through execute_buy). execute_buy succeeds only if the contract holds a fresh ed25519-signed oracle verdict for that mint (allow, or warn plus an explicit human approval); block, missing or expired verdicts fail on-chain. Pay-per-check: request_check reserves a fee from the contract budget and the contract pays the oracle when it posts a verified verdict; cached verdicts are free and every block goes on a shared on-chain blocklist other agents read for free. Execution leg: NEAR Intents 1Click dry quote NEAR to SOL (no Intents testnet, so dry only). Verdicts come from a Solana deployer-reputation check (baseline code, disclosed in README). Tested on NEAR testnet with real reports: $NIHAO (dev sold at 0s, 33% bundled) is blocked, BONK (CAUTION) needs the owner to approve, and a prompt-injected agent's 9 attacks (direct buy, wrong receiver, over-limit, plain transfer, add key, forged verdict...) all fail. 15 contract unit tests. Limits: single oracle key (next: threshold or TEE-attested oracle), testnet only, scripted agent.