Cilium v1.18.15 - Endpoint GC, Backend Quarantine and ip6tables Fixes


Cilium v1.18.15 is a patch release on the 1.18 maintenance branch, published October 15, 2026. It carries three datapath bugfixes that touch endpoint lifecycle, iptables reconciliation, and backend restoration during agent restart. Everything else is dependency maintenance and CI plumbing.

The full release notes and downloads are on the GitHub release page.

The most consequential fix concerns endpoint identity reuse. During a garbage collection iteration, the agent could delete a newly created endpoint because it reused the ID of an endpoint that had already been GC’d. The fix went in through backport PR cilium/cilium#49121 and upstream PR #48905.

If you have seen endpoint flapping or surprising policy drops after node churn, this is a plausible cause. The bug requires ID reuse to trigger, so busy clusters with frequent pod restarts are the most exposed. On a patch release like this, that single fix is worth the upgrade on its own.

A related fix in node_neighbor_discovery corrects the ID used in a delete query (cilium/cilium#48872), so the neighbor table now cleans up the right row.

The iptables reconciliation loop was silently ignoring ip6tables chain rename errors. That means a failed rename left the chain in an unexpected state while the agent believed reconciliation had succeeded. Upstream PR #49024 fixes this, and the fix shipped via backport PR #49333.

For IPv6 enabled clusters, silent failures in rule reconciliation are hard to detect from the outside. Traffic still flows until something restarts a node or reloads rules, then the drift shows up. The fix turns a quiet error into a visible one, which is the right direction for anyone who has debugged asymmetric IPv6 policy behavior.

The third bugfix addresses load balancing backend state. If a backend was in a terminating state when the agent went down, and it became active again while the agent restarted, the restored backend was incorrectly kept as quarantined. Traffic stopped reaching it even though it was healthy again.

The fix comes from @joamaki via upstream PR #48763, backported in cilium/cilium#48795. Operators running rolling agent restarts on clusters with in flight backend lifecycle changes should benefit. The quarantine state is useful when a backend is genuinely failing, but persisting it across a restart against a recovered backend is wrong, and this release corrects that.

The release bumps etcd and golang.org/x/crypto, and pulls in golang.org/x/net v0.60.0 and an OTLP trace exporter update, both flagged as security updates in the notes. If you track CVE exposure in your image provenance, these are the changes that reduce it.

The Envoy sidecar images moved several times during the window, ending at v1.38.6-1791537840-00e749f11f69334c625e53fc1ce2fc6c7f34a51f. The Go build image is now on the 1.26.9 digest. None of these require config changes, they land with the new image.

One operational note: the connectivity check manifest moved echo-b to the lower NodePort band (cilium/cilium#48926). If your firewall rules reference the old band for connectivity check pods, you will want to update them.

The bulk of the changelog by line count is CI maintenance: OIDC support for workflows, bigger runners for Envoy conformance, retry logic for stalled kind image pulls, runner variables named after their Ubuntu release, and cleanup of disk space in the documentation workflow. This is the project keeping its own pipeline stable. It does not affect your clusters, and the only reason to care is if you maintain a fork with its own CI, in which case cherry picking a few of these is reasonable.

Container images are published to quay.io/cilium/, including cilium:v1.18.15 with digest sha256:ab7ccf148b235887a7824313ddfa3edc10d76fcfb13a38e4d5dd213b08533f0a. Pin the digest in your manifests rather than the floating tag. This is a bugfix release with no breaking changes in the notes, so the upgrade is a standard patch rollout, with the endpoint GC fix being the reason not to sit on v1.18.14.