Skip to content

Destynova2/rfc-hit

Repository files navigation

Human Intent Token (HIT)

Status: Work In Progress — This is an early draft. Everything is subject to change. Feedback welcome via Issues.

HIT is a cryptographic protocol that ensures every tool action performed by an AI agent can be traced back to an authenticated human decision. It adds three missing layers to agentic AI systems: signed human intent, independent multi-agent validation, and tamper-proof audit trails.

The Problem

When an AI agent calls a tool (edit a file, run a command, make a payment), nothing cryptographically proves a human requested that action. The agent could be hallucinating, prompt-injected, or acting on stale instructions. As agents gain access to cloud infrastructure, databases, and financial systems, this becomes an existential risk.

How It Works

Human ──sign──> HIT (JWT) ──> Agent ──> HIT Gateway ──> Tool
                                         │
                                   1. Verify signature
                                   2. Check scope & TTL
                                   3. Compute risk score
                                   4. Route:
                                      low risk  → auto-approve
                                      mid risk  → AI quorum vote
                                      high risk → human confirm
                                   5. Execute → receipt chain

A HIT is a signed JWT that says: "this agent can use THESE tools, on THESE files, for X hours, with a $Y budget". Every action goes through a gateway that verifies the signature, assesses risk, and records a hash-chained receipt.

Documents

Document Audience Description
HIT-explained.md Everyone 5-minute explainer with analogies and examples
RFC-0001 Implementers Full technical specification (~900 lines)
examples/ Implementers Policy profile templates (grocery, trading, infra, DeFi)
docs/ Implementers Tutorials, reference, how-to guides

Key Features

  • Human attestation — Every action starts with a cryptographically signed human intent
  • Graduated trust — Low-risk auto-approves, high-risk escalates to human
  • Multi-agent quorum — Independent AIs from different providers vote on medium-risk actions
  • Domain scopes — Financial, trading, infrastructure, and DeFi constraints with field-level controls
  • Multi-human co-signing — M-of-N humans with role-based authorization for high-value operations
  • Tamper-proof receipts — Hash-chained execution records for forensic audit
  • Credential injection — Agent never sees API keys or secrets
  • Circuit breakers — Automatic scope revocation when velocity limits are breached
  • MCP-native — Designed as middleware for the Model Context Protocol
  • Self-sovereign — Local keys, no cloud IdP required

Prior Art

HIT is positioned alongside existing work in AI agent authorization:

  • Agentic JWT (IETF) — Closest prior art. HIT adopts agent_checksum, adds tool-level scoping and quorum
  • Google AP2 — Payment-specific agent mandates (60+ companies)
  • Visa TAP — HTTP signatures for agent commerce
  • NIST AI Agent Standards — Federal initiative validating the problem space
  • IETF OAuth WG — 5+ active drafts on agent delegation. HIT operates below OAuth (tool-call level)

Contributing

This is an early-stage RFC. If you have feedback, questions, or want to discuss the approach, please open an issue.

License

TBD — Considering Apache 2.0 or CC BY 4.0 for maximum adoption.

About

RFC-0001: Human Intent Token (HIT) — Authenticated Tool Authorization for Agentic Systems

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

 
 
 

Contributors