The problem

Hosted agent platforms hold master keys and can read prompts and responses — a single point of total failure.

What it does — AI Pipeline track

BlindRelay is a private agent with encrypted inputs.

  1. The master key is split with Shamir's Secret Sharing 3-of-5 over GF(256), implemented from scratch in lib/shamir.js. Shares go to the user's machine, two co-signers, a paper backup, and an offline backup. The whole key is wiped; only a SHA-256 fingerprint remains.

  2. Reconstruction is local and momentary: any 3 shares reconstruct, verified by the fingerprint, with a zeroed buffer. The host gets zero shares.

  3. Everything the host stores is AES-256-GCM ciphertext under an HKDF-SHA-256 session key. The host audit at /host-view shows masterKey: null, keyShares: 0, plaintextPrompts: 0, plaintextResponses: 0.

  4. The LLM call bypasses the host, going direct to an OpenAI-shaped /v1/chat/completions endpoint. A mock provider ships; a real endpoint is configured via environment variable.

Demo

Five separate OS processes on loopback: local agent + UI, untrusted host, LLM stand-in, and two co-signers. Live attacks — 2-share reconstruction and ciphertext tampering — both fail detectably.

Run: node start.js → http://127.0.0.1:8765, node demo-cli.js, node test.js (all passing).

Math and security

Lagrange interpolation at x=0; an information-theoretic guarantee for k−1 shares (zero Shannon information); availability from any 3 of 5 shares; AES-GCM integrity; HKDF per-run keys.

Honest limits

The LLM provider necessarily sees the prompt; k colluding holders can reconstruct; the demo co-signer uses a shared demo token on loopback (production would use mTLS); user-machine compromise defeats client-side schemes; traffic metadata remains visible.

Built with

Node.js (zero dependencies, built-in crypto), Shamir GF(256) from scratch, AES-256-GCM, HKDF-SHA-256, SHA-256, OpenAI-shaped API (mock included), loopback web UI.

Evidence in the repo

docs/demo-run-output.txt, screenshots/terminal-demo.png, test.js (9 test groups passing).

Try it

Build Disclosure: AI-built. Built by AI agents (Tree / Muse fleet) directed by Rene Turcios during the hackathon period; Rene set the direction and reviewed; agents designed, coded, and tested.

Built With

Share this project:

Updates

Submission history