Sentinel

Sentinel is an AI payment security layer that sits between any AI agent and its wallet on NEAR. AI agents holding crypto can be tricked into sending funds via prompt injection, phishing, or compromised instructions. Sentinel prevents this by gating every payment through a policy engine and counterparty risk scorer before the agent signs. Before an agent sends funds, Sentinel checks: spending limits, allowed recipients, daily caps, and human approval thresholds. It then scores the receiver using on-chain data (account age, contract status, balance) plus AI risk analysis. The result is allow, warn, or block. On block, the agent cannot sign — it uses function-call access keys scoped to one contract, never full-access keys. On warn, it pauses for human approval. The demo includes a prompt injection attack: a malicious instruction tries to make the agent send 500 NEAR to a known scammer. Sentinel intercepts at every step — policy violation, counterparty risk, blocklist hit — and blocks the transaction. Full audit trail with TEE attestation proves the policy engine was not tampered with. NEAR enables this through function-call access keys (scoped, revocable, gas-gated) and on-chain account/contract queries for reputation scoring. The shared blocklist creates cross-agent reputation. Everything runs on NEAR testnet with public RPC.