Docker Engine 29.9.0 shipped on 8 October, and one line in its Networking section reads: errors programming the kernel to encrypt the overlay network data plane “could in some circumstances lead to encrypted-overlay-network traffic to some nodes being transmitted in cleartext” 1. The word that stopped me was kernel. By the time that call runs, Docker has already decided the network is encrypted, so the failure is not a typo I can see in a compose file.

The reference count that released a tunnel early

Encrypted overlay traffic to a peer travels through one shared IPsec tunnel per peer. The overlay driver keeps reference counts so a tunnel only comes down when the last endpoint on that peer stops using it. Pull request 53420 says the two sides of that count did not match: addNeighbor() could return without ever calling setupEncryption(), while deleteNeighbor() called removeEncryption() whether or not the setup had worked 2. A peer whose setup failed released a reference it never took. The count could reach zero while an endpoint on the node still used the tunnel, the kernel’s security associations and policy came out from under it, and the VXLAN datagrams for that endpoint went out with no ESP around them.

The same pull request closed the opposite error: tunnel references for remote peers were never released when a sandbox was destroyed, so each leave and rejoin added a reference and the tunnels stayed up. The author calls that direction safe, since a tunnel that never tears down keeps traffic encrypted 2. One bookkeeping error leaked kernel state, the other leaked confidentiality.

The receiver drops it as spoofed

The release note explains why the exposure stays small. “As the receiving peer would drop cleartext packets for encrypted overlay networks as spoofed, the loss of confidentiality is limited to unidirectional flows (e.g. UDP DNS queries) and handshake attempts that never proceed (e.g. TCP SYN)” 1.

That describes the symptom. A two way flow fails, because the reply never arrives. A one way flow carries its payload before anything can answer: the name in a DNS query, the destination and port in a SYN. Those are the bytes that sat uncovered, and a service name in a query tells a reader which internal service exists and which host asked. That maps the stack without completing a connection.

The drop happens at the receiver, filed as a spoofed source 1. The note carries no statement about the sending side’s logs, so that question stays open until I watch a tunnel come down on a node I control.

The label is a string I typed

--opt encrypted is how the feature gets turned on. Docker’s docs say application data among swarm nodes is unencrypted by default, that the flag enables IPsec at the VXLAN level, and that the performance penalty is worth measuring before production 3. The flag is the whole of the interface.

Driver options are a free-form map, though. In 2020 a Stack Overflow question showed --opt encryptedsd accepted at create time, with docker network inspect printing the misspelled key back without complaint 4. The network came up, and the only record of the typo sat in the field an operator reads to confirm encryption. That question is six years old and the shape holds: inspect reports the options I passed, which is a statement of intent. Two networks on one swarm can both carry an option key that looks like the flag, and docker network inspect reports each one the same way, because the map is fixed at creation.

Check one: the options, held against the docs

I list the overlays and read the options map as a string rather than a boolean.

docker network ls --filter driver=overlay --format '{{.Name}}'
docker network inspect <name> --format '{{json .Options}}'

I want the key encrypted with nothing beside it that resembles it. Then I compare those networks against where traffic goes. Docker’s swarm networking page says the ingress network ships unencrypted and has to be customized separately, and published ports reach their tasks through ingress 3. A service can sit on an encrypted overlay, publish a port, and still cross nodes in the clear on the ingress leg. The October release removed a bug. This default is a decision, and it is what I check first.

Check two: ask the kernel

The kernel holds the answer. Docker programs the security associations and policies. ip -s xfrm state lists them with the peer address, the SPI, the algorithm and the counters, and ip xfrm policy shows the policy that steers traffic into them. Root is required.

The comparison uses the Peers list from docker network inspect on a swarm scoped network 3. Each peer on that list should have state on the node that talks to it, and the options map tells me neither. An encrypted overlay with no state is either idle or broken. I read the counters again after a task moves, because a tunnel that exists and has not moved a byte since the reschedule is not the same as one carrying traffic.

Check three: versions on every node

The fix landed in 29.9.0, and the current point release is 29.9.2, dated 20 October 1. Further down the same page, 29.2.1 records that encrypted overlay networks stopped passing traffic to containers on v29.2.0 through v29.0.0, v28.2.2, v25.0.14 and v25.0.13 1. A mixed version swarm loses connectivity across an encrypted overlay, which shows up as an outage rather than a leak. A report from December 2025 describes that symptom, with a service unreachable across two nodes until a fix in 29.1.3 addressed the report 5. I check the engine version on every node hosting a task on an encrypted network, not the manager alone, since no service definition mentions the skew.

What I keep from this

The rule I apply to release notes now: read the bug fix sections for the words kernel, cleartext and encrypted, not only the security section. The entry sits under Networking with no CVE attached, so a feed sorted by score has nothing to show me.

Then I check one node: options map, ip xfrm state, and the peers list beside them. If the state is there, I have an answer about the path that compose cannot give me, and I can re-run the check after a restart.

The label records what I asked for when the network was created. The state on each node answers the question on the day I ask it.


  1. Docker. (2026, October 8). Docker Engine version 29 release notes. Docker Docs. https://docs.docker.com/engine/release-notes/29/ ↩︎ ↩︎ ↩︎ ↩︎ ↩︎

  2. corhere. (2026, August 19). libn/d/overlay: don’t remove in-use IPsec tunnels (Pull request #53420). GitHub. https://github.com/moby/moby/pull/53420 ↩︎ ↩︎

  3. Docker. (n.d.). Manage swarm service networks. Docker Docs. https://docs.docker.com/engine/swarm/networking/ ↩︎ ↩︎ ↩︎

  4. yamass. (2020, February 1). Docker: How to verify if an overlay network is encrypted [Question]. Stack Overflow. https://stackoverflow.com/questions/60017478/docker-how-to-verify-if-an-overlay-network-is-encrypted ↩︎

  5. CristianCapsuna. (2025, December 6). Docker engine v29 encryption on overlay networks broken (Issue #51672). GitHub. https://github.com/moby/moby/issues/51672 ↩︎