Skip to content

Security model

EdgeFetch fetches URLs that come from AI agents, which makes those URLs untrusted input. The main risk is server-side request forgery (SSRF): tricking a worker into reaching your LAN, a cloud metadata endpoint or your Home Assistant.

  • The URL must use https, have a multi-label DNS host, and contain no IP literals, localhost or user:pass@.
  • The resolved IP is checked inside the dialer, right before connect(). This blocks DNS-rebinding tricks where a name resolves to a public IP at check time and a private one at connect time.
  • Blocked ranges include loopback, RFC 1918, CGNAT and Tailscale (100.64.0.0/10), link-local and cloud metadata (169.254.0.0/16), ULA, and IPv6 transition prefixes.
  • Redirects are followed by hand, up to 5 hops, and every hop is checked again.

Headless Chromium is started with every subresource forced through an in-process loopback proxy (--proxy-server=http://127.0.0.1:<port>, --proxy-bypass-list=<-loopback>, --host-resolver-rules=MAP * ~NOTFOUND, and --force-webrtc-ip-handling-policy=disable_non_proxied_udp). Child processes (chromium and pdftotext) run with a scrubbed environment so operator tokens never enter the browser process, and the final post-navigation URL is re-checked against SSRF and robots.txt rules before any DOM is returned.

  • Containers run as non-root 65532:65532, with no-new-privileges and an AppArmor profile in the Home Assistant add-on.
  • On Linux the add-on drops to UID 65532 on every OS thread before opening any socket.
  • One external Go dependency. The hub has zero npm runtime dependencies.
  • Dependabot watches Go modules, npm and GitHub Actions. govulncheck runs in CI.
  • Releases are signed and attested. See Verify a release.

Please report privately through GitHub Security Advisories, not in a public issue. The full threat model is in SECURITY.md.