Skip to main content
Blue Moose logo Blue Moose Field notes
← All articles
docker

Docker Swarm's encrypted overlay can go cleartext

Rows of servers in a data center

I run a single-node setup, so Swarm overlay encryption isn't something I lean on daily. But a Docker advisory published on 10 November made me stop and re-read how that encryption is actually wired up, because the failure mode is the nasty kind: nothing breaks, nothing logs, and your "encrypted" traffic quietly goes out in the clear.

What the advisory says

GHSA-6jhp-7446-hf7f (CVE-2026-115538, rated High, CVSS v4.0 score 7.0) covers Docker Engine versions before 29.9.3. The patched releases are Docker Engine 29.9.3 and, for the Go module, github.com/moby/moby/v2 v2.0.0-beta.29.

Encrypted Swarm overlay networks wrap VXLAN traffic in IPsec (transport mode). That depends on three iptables rules that dockerd installs when a network's first local endpoint is set up: a mangle OUTPUT rule that marks outgoing VXLAN datagrams with the network's VNI so the IPsec policy applies, a filter INPUT rule that drops incoming VXLAN datagrams with that VNI that did not arrive over IPsec, and, in releases with the CVE-2026-92542 fix, a rule that clears marks set by local processes.

With firewalld, dockerd restores its own rules after a reload through firewalld's D-Bus interface. The advisory is blunt that the overlay driver "does not hook into this infrastructure." So when firewalld wipes those rules, nothing puts them back.

Who is actually exposed

All of the following have to be true: the node is in a Swarm with at least one task or container on an encrypted overlay network, and firewalld is running in a way that deletes other applications' iptables rules. That means FirewallBackend=iptables, or the default nftables backend with either a firewalld build lacking upstream commit b7faa74 or any direct configuration (direct rules, chains, passthrough rules).

The triggers are boring, which is the problem:

  • firewall-cmd --reload or --complete-reload, including runs from configuration management
  • stopping firewalld with CleanupOnExit=yes (the default)
  • starting firewalld, including restarts during package upgrades, or a start after dockerd

An Ansible run that reloads firewalld after adding a port is enough.

The impact is two-sided

Until the rules return, the node sends VXLAN traffic for each affected encrypted network in cleartext, so anyone on the network path can read container traffic. It also accepts cleartext VXLAN datagrams with that VNI, so someone who can send UDP to the VXLAN port (default 4789) can inject arbitrary Ethernet frames, the same class of problem as CVE-2023-28840.

The advisory also notes that dockerd does not log the loss. Oddly, two affected nodes keep talking to each other happily in cleartext, while a node that still has its rules drops cleartext from an affected peer, so you may see mysterious connectivity failures in a mixed state. The condition lasts until the network's last local endpoint is removed and another is added, or dockerd restarts, and every further firewalld event repeats it.

What I'd do

Upgrade to 29.9.3. If you can't yet, the advisory lists workarounds: use the nftables backend with a firewalld build containing b7faa74 (firewalld 2.0.0+ or the listed RHEL backports) and remove all direct configuration; restart dockerd after every firewalld reload, stop, or start (which also restarts that node's Swarm tasks); or block UDP 4789 at the internet boundary. That last one stops outside injection but does nothing for confidentiality between your own nodes.

My takeaway is a general one: a security control that lives in firewall rules owned by a different daemon is only as durable as the weakest reload path. If you use encrypted overlays, treat any firewalld change as a Docker event, and find out from your own packet captures, not from the absence of errors, that traffic between nodes is still ESP.

BM
Blue Moose
The moose behind Blue Moose. Full-stack PHP developer — Drupal by day, Symfony by night, tests always.