Loops HouseHushmark

NEAR HACKATHON
Links
Team
1 member- EGOwner
Egwuatu Rita
Overview
The problem: AI coding agents now read pull requests, docs and rules files (AGENTS.md, .cursorrules) and follow the instructions inside them, including text human reviewers never see. One hidden HTML comment or a few invisible characters can tell an agent to approve a change, skip the tests or keep quiet, and once merged it steers every agent session on the team.
What I built: Hushmark, a GitHub check that stops hidden instructions aimed at AI agents before they are merged. Deterministic detectors scan the whole diff, title, description and agent rules files; nothing in the pull request is ever executed. Only the flagged excerpts, never the code base, are reviewed by a model, and a finding is shown only if that review is cryptographically verified. In block mode a finding fails the check and stops the merge. It runs as a GitHub Action, as a GitHub App (live on a demo repository, with /hushmark re-runs for maintainers) and as an MCP server, so AI agents can check what they are about to read. Every review leaves a receipt of hashes only that anyone can verify. 281 tests; real TEE evidence replays offline with make verify.
How NEAR fits: the review runs on NEAR AI Cloud, inside a TEE (Intel TDX with NVIDIA confidential GPUs), so private code is judged without being exposed. For every review Hushmark checks the TEE's attestation with a fresh nonce, Intel's quote, NVIDIA's signed GPU verdict and the TEE's signature over the exact request and response; if any check fails, no finding is shown. Each receipt's hash is anchored by a NEAR smart contract (anchor.hushmark.testnet), written with a function-call key that can do nothing else, so a verdict can never be swapped later and anyone can check it without trusting me. Only two hashes go on chain, by design, to keep private code private.
Demo: tryhushmark.vercel.app