Skip to content

Deployment Hardening

RAVEN opens up to four listeners: BMP for your routers, the HTTP API for the CLI, the Prometheus metrics endpoint and, when you enable it, pprof. None of them authenticates its clients, so who can reach each port is your access control. This page explains what each listener exposes and how to restrict it.

Start by checking what RAVEN actually bound. At startup it logs one listeners bound line with the resolved address of each listener. With the defaults and no allowed-routers, you see:

{"time":"2026-10-18T22:19:48.267905655-04:00","level":"WARN","msg":"BMP listener accepts connections from any source on a non-loopback address; set bmp.allowed-routers to the routers' addresses","subsystem":"bmp","addr":"[::]:11019"}
{"time":"2026-10-18T22:19:48.268150928-04:00","level":"INFO","msg":"listeners bound","bmp":"[::]:11019","api":"127.0.0.1:11020","metrics":"[::]:9595"}

An address of [::] or 0.0.0.0 means every interface. Here BMP and metrics are open to anything that can route to the host, and the API is loopback only.

Restrict BMP sources

Anything that can open a TCP connection to the BMP port can feed routes into your route table, change every posture and metric RAVEN reports, and make it parse attacker-chosen input. An internet scanner found an exposed BMP port on a production RAVEN deployment within a day. Restrict BMP in two places.

1. Set bmp.allowed-routers to your routers' BMP source addresses:

bmp:
  listen: ":11019"
  allowed-routers:
    - 192.0.2.10              # edge1
    - 192.0.2.11              # edge2
    - 2001:db8:100::/64       # core routers' loopbacks

RAVEN closes a connection from any other source at accept, before it reads anything or starts a TLS handshake, logs BMP connection rejected: source not in bmp.allowed-routers at WARN, and counts it in raven_bmp_rejected_connections_total{reason="not_allowed"}. Alert on that counter: see the RAVENBMPRejectedConnection rule in Observability & Metrics.

Use the address the router sources BMP from, which is often its loopback rather than the interface facing RAVEN. If the router's sessions do not appear after you set the list, look for the rejection WARN: it names the address the router connected from.

With allowed-routers empty, RAVEN accepts any source and logs a WARN at startup when the BMP listener is not on loopback. Treat that WARN as a finding.

2. Filter the port in the host firewall as well. The allowlist protects RAVEN; the firewall keeps the port closed to everyone else, keeps connection attempts out of RAVEN's logs, and still holds if someone empties the list in a config change. With nftables, for example:

table inet raven {
  chain input {
    type filter hook input priority 0; policy accept;
    tcp dport 11019 ip saddr { 192.0.2.10, 192.0.2.11 } accept
    tcp dport 11019 ip6 saddr 2001:db8:100::/64 accept
    tcp dport 11019 drop
  }
}

If your routers support it, BMP over TLS with a client CA adds router authentication on top.

Keep the API on loopback

The HTTP API serves the CLI: the full route table, every peer, and the what-if, audit and Flowspec endpoints, including POST /api/v1/flowspec/rules/<key>/toggle-dryrun, which switches a Flowspec rule from dry-run to live injection. It has no authentication. By default it binds to 127.0.0.1:11020, so only users with a shell on the monitoring host can use it.

Keep that default and run the CLI on the monitoring host. If you need the CLI elsewhere, prefer an SSH tunnel to opening the port. Open the tunnel in the background, then run the CLI against your local end of it:

ssh -f -N -L 11020:127.0.0.1:11020 raven-host.example.net
raven routes --posture origin-invalid

If you do set api.listen to a routable address, bind it to a management interface and firewall it to the hosts that run the CLI, as for BMP above. See api.listen.

Decide who may scrape metrics

The metrics endpoint (outputs.prometheus.listen, default :9595) binds to every interface. Metrics carry your routers' names and addresses, every BGP peer's address, your RPKI caches' addresses, and route counts by posture, which tells a reader how many of your routes fail validation. That is reconnaissance value, so decide who needs it:

  • If Prometheus runs on the same host, bind metrics to loopback: listen: "127.0.0.1:9595".
  • Otherwise bind it to the management interface and firewall it to your Prometheus servers.

pprof stays on loopback

The pprof endpoint (debug.pprof.listen) is off by default. When you enable it for a memory investigation, bind it to loopback, for example localhost:6060, and reach it through an SSH tunnel. Profiles reveal process memory contents, and the endpoint is unauthenticated. RAVEN does not stop you binding it elsewhere, but it logs a WARN (pprof endpoint is exposed beyond localhost) when the bound address is not loopback. Turn pprof off again when you are done. See Memory Profiling.

Checklist

Listener Setting Default Recommended
BMP bmp.listen, bmp.allowed-routers :11019, any source Your routers only, by allowlist and host firewall
API api.listen 127.0.0.1:11020 Loopback; SSH tunnel for remote CLI use
Metrics outputs.prometheus.listen :9595 Loopback or management interface, firewalled to Prometheus
pprof debug.pprof.listen off Off; loopback only while profiling