Loops HouseAgentFuzz

NEAR HACKATHON
Links
Team
1 member- RIOwner
Richard Ting Li Zeng
Overview
AgentFuzz is a red-team harness for NEAR agent-wallet guards. Teams give AI agents money through a guard contract and a function-call key, but almost none test those guards adversarially. AgentFuzz attacks them on two layers: LLM-written prompt injections and tool-output poisoning fed to a reference agent that holds only a function-call key, and property-based fuzzing of the contract (boundary amounts, window rollover, look-alike receivers, same-block callback races, key and gas-allowance abuse). Six invariants are checked against the chain's own receipts; every break is shrunk to a minimal, replayable repro. The real contract wasm runs in nearsim, a deterministic in-process NEAR runtime, and is confirmed on a near-workspaces sandbox. Results become a scorecard whose SHA-256 is anchored in a testnet scoreboard contract; agentfuzz verify re-runs it and recomputes the hash. All 8 planted bugs are found; the reference guard holds all 6 invariants. A static web playground runs the same engine in the browser: fuzz any guard build or your own wasm, and run the prompt corpus, with results identical to the CLI. For the agent economy, the scoreboard also records an audit per guard build keyed by its code hash, so before an agent pays or delegates to another agent it can check that wallet's exact build on-chain (agentfuzz check ), and the corpus includes agent-to-agent attacks. For every NEAR user, the site opens with "Is my NEAR wallet safe?": an account safety scan, in plain words with a traffic light, that reads any mainnet or testnet account and flags risky access keys, look-alike names and audited-but-broken contracts (agentfuzz scan ), and "simulate before you sign" runs any transaction's real contract code on a fork of mainnet or testnet in the browser and explains what it would do (agentfuzz simulate).