Loops HouseUnheld

NEAR HACKATHON
Links
Team
1 member- N1Owner
n1c0la5-lab
Overview
Unheld lets an AI agent pay on NEAR without ever holding a key.
An ICP canister owns a NEAR account through threshold Ed25519: the key is split across the subnet's nodes and never exists in one place. The agent, running on NEAR AI Cloud (DeepSeek-V4.1-Flash, TEE attestation and response signatures verified), can only ask the canister for a payment. Before it signs, the canister checks a policy the agent cannot change: a per-payment limit, a daily limit and a receiver allowlist. NEP-413 messages get their own recipient allowlist, and intents.near is always refused. Every request, signed or refused, goes into the canister's audit log.
Attack demo: the agent reads documents with injected change notices (a look-alike receiver, an amount over the limit, payments split against the daily limit, a NEP-413 message for intents.near). Over 70 runs the model was fooled in 4 of 5 attack types; the canister refused every attempt, and forbidden signatures were 0. The verdict is recomputed from the audit log and the policy, not taken from what the model says. Both controls were signed 10 of 10.
The same canister runs on ICP mainnet (ch3s7-hiaaa-aaaao-qqm3q-cai) with the production key key_1 and signed a transfer that NEAR testnet finalized.
Try it: ./setup.sh builds the canister on a local ICP network and replays the attack requests, with no key, no model and no NEAR account. It ends with PASS or FAIL.
Limits, stated in the README: a fooled agent can still spend inside the policy, so the damage is bounded, not zero. Not run on NEAR mainnet.
The code was written with Claude Code.