AI agents that can spend crypto are one prompt-injection away from draining a wallet. AgentFence is an on-chain policy purse on NEAR plus an agent-side wrapper: the human sets per-tx and daily limits, allow/block lists, and a shared reputation feed; the agent receives only a function-call access key scoped to AgentFence (it cannot Transfer NEAR or AddKey). Before every spend the wrapper calls evaluate_receiver (Allow/Warn/Block with reasons). pay() re-checks on-chain — Block always refuses, Warn needs a human. The demo shows a legitimate payment succeeding and a prompt-injected drain attempt being refused without signing. Testnet-ready WASM and 6/6 unit tests ship in the repo. Tracks: AI Security & Defense and Agent Economy.
Loops House