Skip to content

Security hardening

ReD-Pot is built to be attacked. The design goal is that a fully hijacked decoy gains nothing: no route to the manager, no secrets worth having, no way to speak for another sensor.

Built-in protections

Containers

  • Run as uid 10001 with no capabilities (cap_drop: ALL), no-new-privileges, the default seccomp profile, a read-only root filesystem (only tmpfs and data volumes are writable), memory and process limits, log rotation, and no Docker socket.
  • The image has no setuid or setgid files.
  • Host-network decoys start as root with only NET_BIND_SERVICE, SETUID, SETGID and SETPCAP under no-new-privileges. The entrypoint drops at once to uid 10001 and keeps only NET_BIND_SERVICE. A service started as root without that drop refuses to run.

Network

  • On each decoy interface only the decoy's ports answer; everything else looks closed, IPv6 included.
  • Admin ports answer only on the management interface, to their allow-lists.
  • Decoys cannot open the web UI, 127.0.0.1, the LAN or the internet (egress guard). The server does not route for a decoy LAN.
  • See Ports and firewall.

Hijacked decoy

A decoy holds only its own key and a hashed honeytoken index. It cannot open the web UI, cannot rewrite its configuration (read-only mounts) and cannot report as another sensor. The domain controller decoy additionally holds the Kerberos keys of the bait accounts valid on it: password-equivalent for the fake domain only, never the plaintext.

Untrusted input

  • Control characters are stripped and fields capped at the sensor.
  • Every value in the console is escaped, with a strict Content Security Policy.
  • Every SQL query is parameterised; file paths are confined; archive extraction refuses .., links and devices.
  • Protocol parsers (BER, DNS) are bounded and fuzz-tested.
  • Nothing an attacker sends is ever executed. URLs typed into the fake shell are recorded, never fetched.

Floods

One connection budget per decoy, shared by every TCP service: 250 concurrent connections, 20 per source IP, and no new client while the decoy runs 320 threads. SMB sessions without a named logon close after 60 seconds idle. UDP is rate-limited per source and replies are capped (no amplification). Request bodies are capped. A listener that dies is restarted and reported.

Manager

  • TLS 1.2+, HMAC-signed HttpOnly, SameSite=Strict session cookies, throttled login, a CSRF header on state changes and on export, and a separate read-only cookie for wall screens.
  • No outbound connections for alerting: alerts stay in alerts.json and the database.
  • Tried passwords are stored masked.

Your checklist

  • Put decoys on their own interfaces where possible, so attackers on the decoy LAN never see an admin port.
  • Keep the web UI, sensor API and SSH allow-lists as narrow as possible (set-access). Avoid 0.0.0.0/0 when a decoy shares the management interface.
  • Keep the admin token, rpot.json, INSTALL-INFO.txt, exports and backups secret and root-only. Rotate the token with reset-token if it may have leaked.
  • Never put a sensor join string on a command line; use the prompt, --join - or --join-file.
  • Give the manager, decoy interfaces and sensors static addresses (or DHCP reservations) and working NTP.
  • On remote sensors, add a switch or firewall ACL so the decoy IP may only answer connections, never initiate them.
  • Do not run other services on decoy interfaces; the installer offers to stop them.
  • Use the lab bridge mode only for evaluation: there, decoys can reach the Docker host's bridge IP.
  • Keep ReD-Pot updated (git pull, sudo ./install.sh update) and watch the release notes.
  • Verify downloaded images with gh attestation verify if your policy requires provenance.

Reporting a vulnerability

See the Security policy.