io.github.dir-ai/voyager-net

voyager-net

Voyager's network organ — a safe, read-only, authorized audit of one host or domain you own, so an AI agent can find the falle in your infra and describe the fix.

Voyager penetrates the web (@dir-ai/voyager), the repo (@dir-ai/voyager-repo), and — here — networks. This is the sense: it finds problems. Applying fixes ("the hands") is a separate, consent-gated organ, by design — an AI must never mutate live infrastructure on its own.

voyager-net scan yourhost.example.com --authorized
voyager-net scan 10.0.5.20 --authorized --ports 22,80,443,5432

What it checks (read-only)

• Resolve-once + IP pinning (every probe, HTTP included) — a domain is resolved a single time and the DNS, port, TLS and HTTP probes are all pinned to that one IP via a custom socket lookup, so nothing re-resolves the name mid-scan (closes the HTTP DNS-rebinding path). The resolved IP is classified canonically (via ipaddr.js) and any non-public address — loopback, private, link-local, CGNAT, NAT64, IPv4-mapped IPv6, metadata — is refused.
• DNS — A/AAAA/MX/TXT/CAA + email/CA hygiene, and it grades the content: SPF +all (anyone may send) and DMARC p=none (no enforcement) are findings, not just presence checks. All target-controlled records are injection-framed.
• Ports — a bounded common-service probe by plain TCP connect() with real states (open / closed / filtered / unreachable — an OS timeout reads as filtered, not closed), no SYN tricks, no payloads, no range/CIDR sweeps.
• Service fingerprinting — reads the banner a service volunteers on connect (SSH, SMTP, FTP, POP3, IMAP) to identify product + version (framed).
• Unauthenticated-service detection — for auth-expecting services (Redis, Memcached, Elasticsearch, CouchDB, Docker API, ZooKeeper) it sends ONE benign, read-only protocol hello (e.g. Redis PING) and reports the service as exposed only if it answers without asking for credentials. Honest scope: this and the TLS/HTTP inspectors DO send bytes (a ClientHello, a GET /, a protocol hello) — the audit is active-but-safe (no writes, no mutation, no exploit), not purely passive, and it runs only against the host you authorized.
• TLS — the full set of accepted protocol versions (not just the negotiated one), chain trust against the system store (any validation failure is reported, not only self-signed), RSA key size (EC/EdDSA exempted), certificate expiry/issuer.
• HTTP hygiene (graded) — status, Server banner, HSTS/CSP quality (weak max-age, unsafe-inline/wildcard — not just presence), clickjacking protection, per-cookie Secure/HttpOnly (each cookie evaluated on its own), CORS wildcard and the dangerous credentials-with-origin case, version-leak, and HTTP→HTTPS redirect. All pinned to the vetted IP.
• Honest partial — an open TLS/HTTP port that couldn't be inspected is reported as UNKNOWN, never folded into a "no issues" verdict.

Findings that name a detected service+version suggest checking it against a CVE feed — voyager-net detects the version; CVE matching stays a lookup, and CVE probing (Nuclei &c.) is out of scope — that is active testing, a separate, more-gated capability, not a read-only sense.

Each finding carries a severity, confidence, and a described fix — e.g. "certificate expires in 6d → renew and automate ACME", "mysql reachable publicly → restrict to a private network", "no DMARC → publish v=DMARC1; p=quarantine".

Safety — non-negotiable

• Authorized-only, fail-closed. Without --authorized (CLI) / authorized:true (MCP) it refuses. You assert you own / may test the target.
• One host or domain only. CIDR ranges, IP ranges, lists, wildcards, and URLs are rejected — it can never become a mass scanner.
• Cloud metadata endpoints are hard-blocked (169.254.169.254 etc.).
• Read-only. It never applies a fix, exploits, floods, or mutates anything. Bounded concurrency and timeouts; a plain TCP connect, not a scanner's SYN sweep.
• Untrusted output framed. Every target-controlled string — banners, certificate subject/issuer/SANs, Server/CSP headers, CORS value, and DNS TXT/CAA/MX records — is injection-stripped before your model sees it.
• Exit codes: 0 clean · 1 high/critical finding(s) · 2 tool error / not authorized.

Scanning infrastructure you do not own or have explicit permission to test may be illegal in your jurisdiction. This tool is for auditing your own systems.

MCP

voyager-net mcp

Tool: scan_host — same audit, fail-closed (authorized defaults off).

Library

import { scan } from '@dir-ai/voyager-net'
const brief = await scan('yourhost.example.com', { authorized: true })
const urgent = brief.findings.filter((f) => f.severity === 'high' || f.severity === 'critical')

Roadmap

Wrap Prowler/Steampipe (cloud config), Trivy (CVE), Hubble (flow) under the same trust contract; attack-path correlation; drift (IaC declared ↔ actual). Then — separately and consent-gated — the remediation "hands".

License

MIT