Loops HouseArena

NEAR HACKATHON
Links
Team
1 member- DAOwner
Dami
Overview
Your AI agent will pay your bills. A guard is the NEAR contract that limits what it can pay: which payees, how much a day. You probably test your guard with your own tests, and they share its blind spots.
Arena attacks the guard for you. Run ./arena check your_guard.wasm and Arena deploys it on a throwaway local NEAR sandbox, then runs its attack library: two pay calls for one bill in the same block, cap races, outsider calls, invoice emails written to fool the agent. A learning attacker can also read the guard's code first and aim at payments the guard allows but the owner never wanted. Arena scores each attack from final chain balances, not from what the guard reports.
The library grows. A hunter reads new research papers every day and proposes attacks from them. An attack joins the library after a gate shows it holds on a known-good guard and breaks a copy with one planted flaw, 3 runs each: 33 of 54 so far, plus 14 hidden attacks published as a fingerprint.
Results: a planted-flaw guard paid one 100-token invoice twice on NEAR testnet, 5 runs of 5. Proofs on four other teams' guards checked 38 money claims and found one that fails: a refund after UTC midnight let a day pay over its cap.
NEAR's part: guards run as NEAR contracts on near-sandbox. NEAR function-call access keys stop some attacks before the guard runs, and Arena reports those separately. Arena also checks guards that pay through NEAR Intents and Chain Signatures. Results go on NEAR testnet, so an agent can look a guard up before it trusts it. The hunter reads papers through NEAR AI Cloud private inference and saves the attestation.