Loops HouseProving Ground

NEAR HACKATHON
Links
Team
1 member- SAOwner
Sambit Sargam Ekalabya
Overview
AI agents that hold money get prompt-injected: Freysa lost $47K to one message. Rules inside the agent's own code don't help, because a jailbroken model runs that code.
Proving Ground treats the model as untrusted and binds every agent payment to what the human signed, enforced by a NEAR contract. The owner creates a mandate in our vault contract from their own wallet: allowed payees, per-payment, total and daily caps, expiry, and a salted hash of the purpose. The agent only holds a function call key that can call request_payment and nothing else, so it can never transfer funds itself. The contract checks every request deterministically and blocks any rule failure with a reason code. If the rules pass, a veto-only checker running in a TEE on OutLayer (yield/resume) reads the invoice, sealed with X25519 and ChaCha20-Poly1305 so only hashes go on chain, and compares it to the purpose. It can block or park a payment for the owner, never approve one the rules refused.
Live on NEAR testnet with the vault locked (no access keys). A public arena lets anyone attack the real agent. An npm CLI runs 25 attacks against any agent: 25 stopped, 0 funds lost. Technical walkthrough: https://youtu.be/PA7tC3oSsTo