zero-trust-network-mesh

Implement Zero Trust networking for multi-agent swarms. mTLS mutual authentication, egress proxy whitelisting, DNS-over-HTTPS pinning, Ed25519 message signing, and network segmentation for agent clust

SKILL.md

GitHub

Zero Trust Network Mesh

"Never trust, always verify" applied to inter-agent communication. Every packet is authenticated; every connection is authorized; no implicit trust based on network location.

When to Use

  • Securing communication between distributed agent nodes across hosts
  • Preventing a compromised agent from pivoting to other agents via the network
  • Auditing and controlling all outbound API traffic from agent processes
  • Building a production-grade swarm where network isolation is critical

Do NOT use for

  • Single-host local agent setups where all agents run in the same process
  • Development environments where rapid iteration matters more than isolation

mTLS for Agent-to-Agent Communication

# Generate agent identity certificate (per-agent, ephemeral 24h TTL)
openssl ecparam -name prime256v1 -genkey -noout -out agent-42.key
openssl req -new -key agent-42.key \
  -subj "/CN=agent-42/O=yamtam-swarm" \
  -out agent-42.csr
openssl x509 -req -in agent-42.csr \
  -CA swarm-ca.crt -CAkey swarm-ca.key \
  -days 1 -out agent-42.crt

# Agent connects to Bus with mutual TLS
curl --cert agent-42.crt --key agent-42.key \
     --cacert swarm-ca.crt \
     https://swarm-bus.internal:8443/route

Egress Proxy Whitelist

# nginx egress proxy — only allow approved upstream hosts
server {
  listen 3128;
  resolver 127.0.0.1;

  # Allowlist
  location ~ ^/(api\.anthropic\.com|api\.openai\.com|registry\.npmjs\.org) {
    proxy_pass https://$host$request_uri;
  }

  # Block everything else
  location / {
    return 403 "EGRESS BLOCKED: host not in allowlist";
  }
}

Ed25519 Bus Message Signing

import { generateKeyPairSync, sign, verify } from 'crypto';

// Generate ephemeral agent keypair on startup
const { privateKey, publicKey } = generateKeyPairSync('ed25519');

// Sign outbound message
function signMessage(payload) {
  const data = Buffer.from(JSON.stringify(payload));
  return sign(null, data, privateKey).toString('hex');
}

// Verify inbound message
function verifyMessage(payload, sig, senderPublicKey) {
  const data = Buffer.from(JSON.stringify(payload));
  return verify(null, data, senderPublicKey, Buffer.from(sig, 'hex'));
}

WireGuard Agent Tunnel

# Isolated VPN tunnel between two agent nodes
wg genkey | tee agent-a.key | wg pubkey > agent-a.pub
wg genkey | tee agent-b.key | wg pubkey > agent-b.pub

# /etc/wireguard/wg-swarm.conf on agent-a
[Interface]
Address = 10.99.0.1/24
PrivateKey = <agent-a.key>

[Peer]
PublicKey = <agent-b.pub>
AllowedIPs = 10.99.0.2/32

DNS-over-HTTPS Pinning

// Force all DNS resolution through DoH — prevent DNS spoofing
const DOH_ENDPOINT = 'https://cloudflare-dns.com/dns-query';

async function resolveHost(hostname) {
  const res = await fetch(`${DOH_ENDPOINT}?name=${hostname}&type=A`, {
    headers: { Accept: 'application/dns-json' }
  });
  const data = await res.json();
  return data.Answer?.[0]?.data;
}

Anti-Fake-Pass Checklist

  • mTLS: both client and server present certificates (mutual, not just server-side TLS)
  • Certificate TTL ≤ 24 hours for agent identities
  • Egress proxy tested: curl https://evil.com from agent → 403
  • Ed25519 signature verified before processing any Bus message
  • DNS resolution goes through DoH — not system resolver (/etc/resolv.conf)