Skip to content

Latest commit

 

History

History
72 lines (59 loc) · 2.22 KB

File metadata and controls

72 lines (59 loc) · 2.22 KB

RFC 0002: Agent Fingerprint Specification

Status: Draft
Author: Anna Popivanova
Created: 2025-10-01
Tags: fingerprinting, identity, verification, attestation


Motivation

In a network of autonomous agents, verifying identity is critical. Traditional identifiers (e.g., UUIDs, public keys) are insufficient when agents are mutable, composable, or cloned. This RFC defines a fingerprinting mechanism that captures both static and behavioral traits of an agent — enabling cryptographic and contextual identity assurance.


Goals

  • Generate reproducible fingerprints for agents
  • Support both static (code, weights) and dynamic (behavioral) traits
  • Enable attestation and verification across sessions
  • Resist spoofing, replay, and impersonation

Fingerprint Structure

{
  "agent_id": "airlock:agent:sha256:abc123...",
  "version": "2025.10",
  "fingerprint": {
    "static": {
      "model_hash": "sha256:...",
      "weights_hash": "sha256:...",
      "config_hash": "sha256:..."
    },
    "dynamic": {
      "emoprint_hash": "sha256:...",
      "response_profile": {
        "hedging": 0.72,
        "provocation_reflex": 0.91,
        "humor_signature": "dry-sarcastic",
        "tone_volatility": 0.18
      }
    },
    "environment": {
      "runtime_hash": "sha256:...",
      "platform": "linux-x86_64",
      "dependencies": ["torch==2.1.0", "numpy==1.26.0"]
    }
  },
  "timestamp": "2025-10-14T15:00:00Z",
  "signature": "ed25519:..."
}

Verification Logic

  • Static hashes must match known agent builds
  • Dynamic profile must fall within expected bounds
  • Signature must be verifiable against registry public key
  • Environment hash must match attested runtime (see RFC 0003)

Emoprint

Agents may include an emoprint, capturing affective telemetry from affective tone, sentiment patterns, and challenge-response patterns. See RFC 0008.

Threat Vectors

  • Hash collision spoofing — mitigated via SHA-256 and signed payloads
  • Behavioral mimicry — mitigated via emoprint drift detection
  • Environment forgery — mitigated via attestation (RFC 0003)

Examples

See /examples/fingerprint/agent-v1.json for a full payload.