# Cilium v1.20.5

2026-11-17 · [Original release notes](https://github.com/cilium/cilium/releases/tag/v1.20.5) · [Full summary](https://ratatosk.io/en/releases/cilium/v1.20.5)

Security 5 · Breaking changes 6 · Important fixes 40 · Enhancements 2 · Dependency updates 17 · Other changes 26

## Security (5)

### Egress `toServices` wildcard access fixed

- Affects: egress rules combining `toServices` with `toPorts`

When the referenced Service was absent or selector-less with no backends, these rules allowed traffic to any destination on the specified ports. That traffic is now denied until the Service exists and, if selector-less, has backends.

> * Fix egress policy rules that combine `toServices` with `toPorts` allowing traffic to any destination on those ports when the referenced Service doesn't exist or is selector-less with no backends. (Backport PR cilium/cilium#49638, Upstream PR cilium/cilium#49430, @christarazi)

### Hubble query redaction fixed for unparseable Envoy URLs

- Affects: Hubble setups with `--hubble-redact-http-urlquery` enabled and Envoy access-log URLs that fail to parse
- Additional action: The `--hubble-redact-http-urlquery` setting must be enabled for query redaction to apply.

When an Envoy access-log URL failed to parse, its query was folded into the path and could still appear in the exported Hubble flow URL despite redaction being enabled. The fallback now keeps the query separate so redaction removes it.

> * Fix HTTP URL query not being redacted in Hubble flows when the Envoy access log URL fails to parse. (Backport PR cilium/cilium#49802, Upstream PR cilium/cilium#49427, @sage-mode-hunter)

### [Security status unclear] Restored-endpoint policy rule matching fixed

- Affects: setups with restored endpoints and `enable-policy: always` or `default`

During an agent restart, a policy update could occur while restored endpoints were waiting for identity registration, so only wildcard-subject rules matched. With `enable-policy: always` this caused packet drops; with `default` it tended to fail open. The affected endpoints now also receive applicable CCNP/CNP rules with `matchLabels` endpoint selectors.

> * Fix restored endpoints getting only policy rules that select all endpoints. (Backport PR cilium/cilium#49802, Upstream PR cilium/cilium#49091, @jaredledvina)

### Strict egress encryption CIDR matching fixed

- Affects: strict egress encryption using `--encryption-strict-egress-cidr`

When the CIDR contains host bits, the unmasked address did not match the strict range, so unencrypted traffic that should have been dropped left the node in plaintext without a warning. After the upgrade, the CIDR is masked before BPF uses it, and unencrypted traffic in that range is dropped.

> * Fix strict egress encryption allowing unencrypted traffic when --encryption-strict-egress-cidr contains host bits. (Backport PR cilium/cilium#49558, Upstream PR cilium/cilium#49446, @fds7280)

### [Security status unclear] IPsec overlay traffic stays plaintext for peers without an SPI

- Affects: IPsec clusters using overlay mode (Inferred)
- Additional action: The new flag must be enabled

During a staggered IPsec rollout, traffic from a node with IPsec enabled to a node that has not enabled it could be marked for encryption and dropped by the catch-all XFRM policy. With the flag enabled, that traffic is handled as plaintext overlay traffic until the remote node has IPsec enabled, when encryption marking resumes.

> * ipsec: allow plaintext overlay traffic during migration (Backport PR cilium/cilium#49625, Upstream PR cilium/cilium#49245, @sjohnsonpal)

## Breaking changes (6)

### Datapath-loader peak RSS logging corrected to KiB

- Affects: datapath-loader compilation logs (Inferred)

When the datapath loader logged peak RSS from `getrusage`, it treated `Maxrss` as bytes even though `getrusage` returns KiB. The value is now logged in the `rssKiB` field.

> * datapath/loader: `getrusage` returns KiB not bytes (Backport PR cilium/cilium#49921, Upstream PR cilium/cilium#49769, @andreaterzolo)

### Bandwidth manager prevents out-of-horizon timestamps from starving egress

- Affects: rate-limited pods using the bandwidth manager

When a rate-limited pod received a packet with a departure time beyond the drop horizon—possible with `SO_TXTIME` or TCP pacing after a long RTT sample collapsed the socket's pacing rate—Cilium could drop all its egress traffic for up to tens of seconds. The packet is now dropped in BPF without moving `t_last` beyond the horizon.

> * Fix the bandwidth manager dropping all egress traffic of a rate-limited pod for up to tens of seconds after a single packet carried a departure time beyond the drop horizon. (Backport PR cilium/cilium#49802, Upstream PR cilium/cilium#49653, @IraSkyx)

### `CiliumEndpointSlice` reconciliation checks pod UIDs

- Affects: `CiliumEndpointSlice` controller setups where pods may be recreated under the same name (Inferred)

When a pod is recreated with the same name but a different UID, its stale slice entry is identified and queued for pruning.

> * operator/ces: fix stale CES entry for recreated same-name pod (Backport PR cilium/cilium#49978, Upstream PR cilium/cilium#49824, @smagnani96)

### Duplicate endpoint regeneration log attributes removed

- Affects: setups that emit endpoint regeneration statistics logs (Inferred)

Datapath-realization attributes appeared twice because `stats.datapathRealization.GetMap()` was already included in `stats.GetMap()`. The log now iterates over the combined statistics map once, removing the duplicate attributes.

> * Remove duplicated attributes from endpoint regeneration statistics logging (Backport PR cilium/cilium#49921, Upstream PR cilium/cilium#49855, @fristonio)

### DNS proxy registration failures now terminate agent startup

- Affects: agent startup with endpoint restoration or regeneration and DNS proxy initialization (Inferred)

Endpoint regeneration could create DNS redirects before the proxy registered its listener port, racing with static proxy port setup; a registration failure previously logged a warning and startup continued. Restored endpoints are now held until DNS initialization finishes, and a registration failure is logged as an error and terminates the agent.

Evidence: “If DNS proxy port registration fails, return the error to Hive's job runner, which logs it at error level and terminates the agent. Previously, this failure logged a warning and startup continued.” https://github.com/cilium/cilium/pull/49540

> * Fix agent bootstrap to order endpoint regeneration after DNS proxy initialization to prevent proxy port race. (Backport PR cilium/cilium#49802, Upstream PR cilium/cilium#49540, @jrajahalme)

### `cilium-operator` exits when its controller-runtime manager fails

- Affects: setups running `cilium-operator` with controller-runtime controllers, including Gateway API, `CiliumEnvoyConfig`, or MCS-API service export

Previously, when the manager failed, the controllers stopped while the operator kept running, `/healthz` returned 200, and a leader could keep renewing its lease. Now the container restarts and the lease is released.

Evidence: “Any failure from mgr.Start() is only counted in cilium_operator_hive_jobs_runs_failed. The process keeps running, /healthz keeps returning 200 and a leader elected operator keeps renewing its lease while every controller-runtime controller: Gateway API, CiliumEnvoyConfig, MCS-API service export, is gone, so the standby never takes over and the only recovery is a manual pod delete.” https://github.com/cilium/cilium/pull/48385

> * The cilium-operator now exits when its controller-runtime manager fails, instead of continuing to run with its controllers stopped. (Backport PR cilium/cilium#49802, Upstream PR cilium/cilium#48385, @ryanjwong)

## Important fixes (40)

### Tunnel-traffic `NOTRACK` rules now exclude transit traffic

- Affects: setups using Cilium's iptables rules for tunnel traffic (Inferred)

Previously, destination-port-only matching could let legitimate transit packets bypass the node’s normal packet path. The rules now match `addrtype LOCAL`, so they do not match transit traffic that should be processed as non-local.

> * datapath/iptables: scope iptables rules for tunnel traffic (NOTRACK) to local node addresses by matching addrtype (Backport PR cilium/cilium#50040, Upstream PR cilium/cilium#49266, @bersoare)

### Hubble deleted-pod cleanup for dynamic metrics fixed

- Affects: Hubble dynamic metric configuration

When an endpoint was deleted with dynamic metric configuration, cleanup was gated on the static `EnabledMetrics` collection, leaving the deleted pod’s series behind. Active dynamic metric handlers now clean up those series under the existing one-minute grace period and deletion rules.

> * Fix Hubble dynamic metrics retaining deleted-pod series by applying endpoint-deletion cleanup to the active dynamic metric handlers. (Backport PR cilium/cilium#49802, Upstream PR cilium/cilium#49712, @alexcastilio)

### Cilium-generated ICMP error reply checksums fixed

- Affects: setups where Cilium generates ICMP error replies, including no-backend reply paths

Before this fix, parts of the quoted packet were omitted from the checksum when Cilium generated an ICMP error reply, producing an incorrect checksum.

> * Fix incorrect checksums in ICMP error replies generated by Cilium. (Backport PR cilium/cilium#50040, Upstream PR cilium/cilium#49459, @yckao)

### Endpoint cleanup protects replacement CiliumEndpoints

- Affects: endpoint replacements where both endpoints share a CiliumEndpoint

After a failed sandbox setup, an old local endpoint could remain and its cleanup could remove the replacement's manager references and delete their shared CiliumEndpoint. Cleanup now preserves references belonging to another endpoint and checks CiliumEndpoint ownership and resource version before updating or deleting it.

> * fix incorrect CiliumEndpoint deletion during endpoint replacement (Backport PR cilium/cilium#49854, Upstream PR cilium/cilium#49573, @novikovfred)

### Multi-pool IPAM CIDR release ordering fixed

- Affects: multi-pool IPAM

Each excess-CIDR release call flipped the allocator order, spreading pod IPs across CIDRs instead of filling the oldest first. This kept unused IPs from being released and left nodes holding more CIDRs than needed; the allocator now preserves its order so excess unused CIDRs can be released.

> * Fix multi-pool IPAM spreading pod IPs across multiple CIDRs instead of filling the oldest CIDRs first. This kept unused IPs from being released, so nodes held more CIDRs than they needed. (Backport PR cilium/cilium#49625, Upstream PR cilium/cilium#49418, @41ks)

### False increments in the Hubble ring-buffer lost-events metric fixed

- Affects: Hubble setups using the `hubble_ring_buffer` lost-events metric

The metric used to increment whenever a reader reached the ring-buffer boundary, including expected stops during backward scans where no events were lost. It now increments only when a forward-reading consumer catches an actual lost event.

> * hubble: Fixed false-positive increments of the `hubble_ring_buffer` lost events metric. (Backport PR cilium/cilium#49921, Upstream PR cilium/cilium#46585, @chez-shanpu)

### Cloud IPAM instance-cache resync race fixed

- Affects: Cloud IPAM instance-cache resyncs on AWS, Alibaba Cloud, or Azure

If an update arrived after a full resync fetched its snapshot but before it replaced the cache, the resync could drop a newly attached ENI on AWS or Alibaba Cloud, or restore a deleted instance on any of the three providers. Updates now apply after the resync replaces the cache, preventing those stale-state effects.

> * ipam: prevent full resyncs from overwriting cache updates (Backport PR cilium/cilium#49625, Upstream PR cilium/cilium#49008, @aanm)

### L2 announcer deadlock fixed

- Affects: installations using the L2 announcer

After many selected services were removed at once, the L2 announcer could deadlock while stopping a service's leader election, leaving new IP announcements stopped until the agent restarted.

> * l2announcer: Fix L2 announcements of new IPs stopping until the agent restarts after many selected services are removed at once (Backport PR cilium/cilium#49802, Upstream PR cilium/cilium#49645, @aanm)

### L2 announcer now drops stale leader events

- Affects: installations using the L2 announcer

When a selected service was replaced while its leader event was still in flight, the late event could remove the new service’s entries. Such events are now dropped without recalculating the L2 entries.

> * l2announcer: Fix L2 announcements of new IPs stopping until the agent restarts after many selected services are removed at once (Backport PR cilium/cilium#49802, Upstream PR cilium/cilium#49645, @aanm)

### L2 announcement Lease name collisions fixed

- Affects: installations using L2 announcements

Before this release, two unrelated Services could produce the same Lease name when their namespace and Service name were joined with `-`. New names join them with `.`, which cannot occur in either component, so the names are unambiguous.

> * l2announcer: use unambiguous separator in Lease name to avoid collision (Backport PR cilium/cilium#49921, Upstream PR cilium/cilium#48138, @waterWang)

### BPF session-affinity cleanup fixed

- Affects: BPF load balancing with session affinity

If a service or its endpoints changed before retry after BPF reconciliation encountered an error, session-affinity entries could be deleted unnecessarily or survive a restart. The reconciler now tracks affinity use and prunes affinity matches.

> * loadbalancer: Fixes to cleaning up stale data after BPF reconciliation had encountered an error and the service or endpoints had changed before retry was performed (Backport PR cilium/cilium#49802, Upstream PR cilium/cilium#48919, @joamaki)

### Stale load-balancer quarantine state cleanup fixed

- Affects: load-balancer backend quarantine with restored state

If BPF reconciliation encountered an error and the service or endpoints changed before the retry, restored quarantine state could remain for removed backends or frontends no longer observed. Reconciliation and pruning now discard that stale state.

> * loadbalancer: Fixes to cleaning up stale data after BPF reconciliation had encountered an error and the service or endpoints had changed before retry was performed (Backport PR cilium/cilium#49802, Upstream PR cilium/cilium#48919, @joamaki)

### Correct source identity in no-backend service drop notifications

- Affects: service no-backend response handling

When `nodeport_svc_lb{4,6}()` tail-called into `CILIUM_CALL_IPV{4,6}_NO_SERVICE` without storing the source identity in `CB_SRC_LABEL`, the handler could report a stale or unrelated identity in the drop notification. It now reports the real source identity.

> * Report correct source identity in drop notification for no-backend service traffic. (Backport PR cilium/cilium#49802, Upstream PR cilium/cilium#49544, @tklauser)

### BPF CIDR identity range aligned with allocator limit

- Affects: setups using CIDR identities above `0x100FFFF`

When the allocator assigned a CIDR identity above `0x100FFFF` (up to `0x1FFFFFF`), `identity_is_cidr_range()` and `identity_is_world()` did not recognize it.

> * bpf: fix CIDR_IDENTITY_RANGE_END out of sync with MaxLocalIdentity (Backport PR cilium/cilium#49802, Upstream PR cilium/cilium#49715, @smagnani96)

### Policy computation drops results for removed identities

- Affects: policy computations for local identities

If an identity is removed, or removed and added again, after its policy-computation request is queued, the result is not stored or retried, and its policy is released.

> * policy/compute: Drop results for identities removed mid-computation (Backport PR cilium/cilium#49802, Upstream PR cilium/cilium#49484, @christarazi)

### MCS-API EndpointSlice mirror ownerReferences check fixed

- Affects: Cluster Mesh MCS-API EndpointSlice mirroring (Inferred)

When a derived EndpointSlice had no owner references, the operator could index `OwnerReferences[0]` and crash.

> * clustermesh: fix ownerReferences check in MCS-API EndpointSlice mirror (cilium/cilium#50009, @brb)

### Endpoint BPF regeneration rollback on failure fixed

- Affects: setups that perform endpoint BPF regeneration (Inferred)

When endpoint BPF regeneration returns an error, finalization now uses that final error to resolve rollback state and run the rollback.

> * endpoint: use final regeneration error to resolve rollback state (Backport PR cilium/cilium#49675, Upstream PR cilium/cilium#49471, @jrajahalme)

### L7 redirect IDs now distinguish parser types

- Affects: HTTP and TLS L7 rules on the same endpoint and port

When both rule types are configured on the same endpoint and port, they now use separate proxy listeners.

> * Extend L7 redirect ID defintion with L7 parser type, so that http and tls rules on same endpoint/port use separate proxy listeners. (Backport PR cilium/cilium#49625, Upstream PR cilium/cilium#49549, @nezdolik)

### NodePort load-balancer entry leak fixed

- Affects: NodePort services with different `externalTrafficPolicy` and `internalTrafficPolicy` settings

Deleting these services could leave expanded node-address frontends, reverse NAT entries, service ID allocator entries, and backend references leaked until the agent restarted.

> * Fix a leak of load balancer BPF map entries for NodePort services that have different external and internal traffic policies. (Backport PR cilium/cilium#49802, Upstream PR cilium/cilium#49373, @vt0r)

### NodePort and HostPort source-port reuse timeout fixed

- Affects: NodePort and HostPort connections to local backends

After a connection closed, a client reusing its source port for a direct connection could have the backend’s SYN-ACK reverse-NATed to the old frontend, causing a timeout. The stale reverse NAT is skipped in this case, but direct connections under an L7 ingress policy still time out.

> * Fix connections to a pod timing out when the client reuses the source port of a closed NodePort or HostPort connection to that pod. (Backport PR cilium/cilium#49625, Upstream PR cilium/cilium#47641, @aanm)

### BPF masquerade access to local services restored

- Affects: BPF masquerade with kube-proxy (`KPR` disabled) and `eTP=Local` LoadBalancer services, except in hybrid routing

Previously, pods in this setup could not reach `eTP=Local` LoadBalancer VIPs or NodePorts on their own node because replies bypassed the stack and DNAT was not reversed. The hairpin SNAT rule restores access; backends see the `cilium_host` address, and identity-based policy still applies.

> * Fix pods failing to reach LoadBalancer VIPs and NodePorts of eTP=Local services via kube-proxy with BPF masquerade and KPR disabled. (Backport PR cilium/cilium#49921, Upstream PR cilium/cilium#49066, @maxskokov)

### Consistent `cilium_devices` map definitions with preallocation

- Affects: setups with `--preallocate-bpf-maps` enabled

The agent and loader could replace each other’s pinned `cilium_devices` map, leaving datapath programs using empty maps and causing endpoint regeneration to fail with a `file exists` error. The map definitions now agree, preventing this behavior.

> * Fix the datapath using empty cilium_devices maps, and "file exists" errors on endpoint regeneration, when `--preallocate-bpf-maps` is enabled. (Backport PR cilium/cilium#50040, Upstream PR cilium/cilium#49882, @spnngl)

### Operator ENI multi-pool IP release cap

- Affects: nodes using ENI multi-pool IP allocation

When the operator released unused IPs, it could take a node below `min-allocate`, causing those IPs to be allocated again.

> * Fix the operator releasing ENI multi-pool IPs below min-allocate, causing nodes to immediately allocate the IPs again. (Backport PR cilium/cilium#49480, Upstream PR cilium/cilium#49125, @jaredledvina)

### Shared-backend socket termination fix

- Affects: load-balancer backends sharing an address across services or sources

Removing a backend from one service could terminate its sockets and abort connections even when another service or source still had a live backend at that address. Connections now remain intact while any instance at the address is alive.

> * Fix unintended socket termination when a backend is removed from one service but remains alive and in use by another. (Backport PR cilium/cilium#50040, Upstream PR cilium/cilium#49936, @ysksuzuki)

### DNS proxy rules refresh with incremental policy updates

- Affects: endpoints using DNS egress policy

When a destination’s security identity changed during an incremental policy update, stale DNS proxy-rule snapshots caused requests to newly selected identities to be denied. The refreshed rules prevent those denials.

> * Fixed an issue where DNS egress policy could deny requests after a destination’s security identity changed. DNS proxy rules now refresh alongside incremental policy updates. (Backport PR cilium/cilium#49802, Upstream PR cilium/cilium#49699, @tklauser)

### Operator identity creation for named-port pods fixed

- Affects: pods with named ports when `identity-management-mode` is `operator` or `both`

The operator omitted named-port labels from its identity key while the agent included them, so the agent would not select the created identity and pod creation failed.

> * Fixes a bug that prevents cilium identity creation for pods with named ports when operator-managed-identity was enabled. (Backport PR cilium/cilium#49585, Upstream PR cilium/cilium#48662, @tamilmani1989)

### Gateway API route ordering made deterministic

- Affects: Gateway API routes (Inferred)

During reconciliation, route order could differ between runs; ties now break by creation timestamp, then namespace and name alphabetically.

> * gateway-api: fix an issue when routes order might not be consistent between reconciliations. (Backport PR cilium/cilium#49802, Upstream PR cilium/cilium#49504, @arybolovlev)

### Default Helm tolerations now include the etcd taint

- Affects: Helm installations using the default `operator.tolerations`

In single-node or bootstrap clusters with stacked etcd, an etcd-tainted control-plane node could leave the operator unscheduled. With the chart default, the operator now tolerates the etcd taint and can schedule during cluster initialization.

> * helm: Add node-role.kubernetes.io/etcd to default cilium-operator tolerations (Backport PR cilium/cilium#49802, Upstream PR cilium/cilium#49371, @ankitaojha123)

### Unchanged XFRM policies left in place

- Affects: Cilium IPsec with XFRM policy reconciliation (Inferred)

Previously, each node-reconciler run reapplied every XFRM policy, even unchanged ones; the resulting route-generation bump could invalidate a cached route borrowed by an in-flight packet and drop it as `XfrmInError` or `XfrmFwdHdrError`. The agent now leaves an unchanged installed policy in place, avoiding that packet-drop race.

> * ipsec: Skip no-op XFRM policy updates (Backport PR cilium/cilium#50040, Upstream PR cilium/cilium#49883, @sjohnsonpal)

### Reconciler tracking of NodePort expanded frontends fixed

- Affects: NodePort expanded frontends

When realization of a frontend errors and NodePort addresses change before it is retried, the reconciler could lose track of it. Expanded addresses are now recorded before reconciliation, keeping the frontend tracked through that situation.

> * loadbalancer: Fixes to cleaning up stale data after BPF reconciliation had encountered an error and the service or endpoints had changed before retry was performed (Backport PR cilium/cilium#49802, Upstream PR cilium/cilium#48919, @joamaki)

### Service reflector retries conflicting frontend upserts

- Affects: Kubernetes Services with frontend ownership changes

If a Service requested a frontend before its previous owner released it, the failed upsert left the replacement Service unprogrammed. The reflector now retries when ownership changes or the frontend is removed; tracked conflicts are cleared on Service deletion and rebuilt on full sync.

> * loadbalancer: retry Service upserts blocked by frontend conflicts (Backport PR cilium/cilium#50045, Upstream PR cilium/cilium#49942, @mhofstetter)

### ENI IPAM recovery after full-resync cache drops fixed

- Affects: ENI IPAM

When a full ENI resync dropped a node from the instance cache, it was not recovered through this mechanism and required an operator restart. A subsequent IP pool maintenance pass can now retrigger synchronization once, until it is re-armed or the instance is seen again.

> * operator/ipam: recover ENI IPAM nodes that were dropped from the instance cache by a full ENI resync, instead of requiring an operator restart (Backport PR cilium/cilium#49480, Upstream PR cilium/cilium#46839, @ryanjwong)

### VTEP rejects non-IPv4 mask addresses

- Affects: setups with VTEP enabled and `--vtep-mask` configured

IPv6 and IPv4-mapped IPv6 values previously passed validation, then caused the agent to panic when the VTEP manager wrote the mask into `cilium_vtep_map`. They are now rejected during configuration validation with an error stating that the mask must be an IPv4 address.

> * option: validate vtep-mask is an IPv4 address (Backport PR cilium/cilium#49802, Upstream PR cilium/cilium#49271, @wrzwrz123)

### AWS ENI IP release delay increased by default

- Affects: AWS ENI IPAM using the default `--excess-ip-release-delay`

The agent could reclaim an IP after the operator had started releasing it. The longer default delay lowers the possibility of this happening.

> * pkg/ipam: Mitigate an issue where the agent reclaims an IP once the operator has started releasing it by lowering the possibility of this scenario. The operator now waits 30 minutes to release an IP by default up from 3. (Backport PR cilium/cilium#50047, Upstream PR cilium/cilium#49857, @41ks)

### Real IPsec configuration passed to `nodeHandler`

- Affects: IPsec setups using `--use-cilium-internal-ip-for-ipsec`

`nodeHandler` now applies the behavior configured by `--use-cilium-internal-ip-for-ipsec`.

> * Use the real IPsec config in nodeHandler, which restores the `--use-cilium-internal-ip-for-ipsec` behavior (Backport PR cilium/cilium#49838, Upstream PR cilium/cilium#49827, @vakalapa)

### WireGuard peer update retries after IPCache identity changes

- Affects: WireGuard setups where IPCache identity changes trigger peer updates

A transient error while updating peers during an IPCache identity change could leave kernel `AllowedIPs` out of sync with the IPCache, disrupting traffic for affected IPs over the long term. Updates are retried up to three times with exponential backoff, allowing recovery when a retry succeeds.

> * wireguard: Fixed an issue where transient errors updating WireGuard peers could cause AllowedIPs to remain permanently out-of-sync with the IPCache, resulting in long-term traffic disruption. (Backport PR cilium/cilium#49802, Upstream PR cilium/cilium#48664, @meel-ayush)

### Stale ingress connection-tracking entries recreated on verdict mismatch

- Affects: ingress proxy redirection

A new connection reusing the tuple of an ingress connection-tracking entry left with `proxy_redirect` set by an earlier L7 rule could have its SYN-ACK sent to the ingress proxy, which had no socket for it, leaving the handshake hung.

> * bpf: Recreate ingress CT entry if proxy_redirect is stale (Backport PR cilium/cilium#50040, Upstream PR cilium/cilium#49592, @christarazi)

### Hook program file-descriptor leak fixed

- Affects: plugins that load hook programs

Plugins leaked a file descriptor while loading hook programs.

> * Fix issue where plugins leaked a file descriptor while loading hook programs. (Backport PR cilium/cilium#49625, Upstream PR cilium/cilium#49521, @jrife)

### Bundled SPIRE tolerations now cover etcd-tainted nodes

- Affects: setups using the bundled SPIRE agent on etcd-tainted nodes

Previously, the agent could not run there, preventing a cilium-operator leader on such a node from completing workload identity registration in SPIRE.

> * helm:  allow SPIRE agent to schedule on etcd-tainted nodes. (Backport PR cilium/cilium#49921, Upstream PR cilium/cilium#49667, @weizhoublue)

### Leaked policies from recomputation fixed

- Affects: setups that recompute identity policies and replace existing policies (Inferred)

When policy recomputation replaced an existing policy, the replaced policy could remain reachable afterward, leaking superseded policies.

> * policy/compute: Ensure superseded policies are not leaked (Backport PR cilium/cilium#49870, Upstream PR cilium/cilium#49639, @christarazi)

## Enhancements (2)

- Load-balancer ID allocator metrics added
- Datapath map sweeper adds `DisabledMapsHook` for disabled-map cleanup (To use: `DisabledMapsHook` registration through the `disabled-maps-hooks` group)

## Dependency updates (17)

### `github.com/go-jose/go-jose/v4` updated to v4.1.6

- [CVE-2026-116587](https://www.cve.org/CVERecord?id=CVE-2026-116587), [GHSA-mgcc-cww5-fp2q](https://github.com/advisories/GHSA-mgcc-cww5-fp2q)

Versions through 4.1.5 were vulnerable; 4.1.6 is the patched version.

> * chore(deps): update module github.com/go-jose/go-jose/v4 to v4.1.6 [security] (v1.20) (cilium/cilium#49656, @cilium-renovate[bot])

### `github.com/gopacket/gopacket` updated to `v1.7.2`

- [CVE-2026-119299](https://www.cve.org/CVERecord?id=CVE-2026-119299), [GHSA-358w-w75h-x6rx](https://github.com/advisories/GHSA-358w-w75h-x6rx)

In v1.7.1, crafted SCTP segments could trigger out-of-bounds panics in the SCTP chunk sub-decoder; this issue is fixed in v1.7.2.

> * fix(deps): update module github.com/gopacket/gopacket to v1.7.2 [security] (v1.20) (cilium/cilium#49952, @cilium-renovate[bot])

### `github.com/klauspost/compress` updated to v1.20.2

- [CVE-2026-116591](https://www.cve.org/CVERecord?id=CVE-2026-116591), [GHSA-fpmh-frq8-7rj8](https://github.com/advisories/GHSA-fpmh-frq8-7rj8), [GO-2026-6865](https://osv.dev/vulnerability/GO-2026-6865)

> * fix(deps): update module github.com/klauspost/compress to v1.20.2 [security] (v1.20) (cilium/cilium#49873, @cilium-renovate[bot])

### `golang.org/x/crypto` updated to v0.59.0

> * fix(deps): update module golang.org/x/crypto to v0.59.0 [security] (v1.20) (cilium/cilium#49971, @cilium-renovate[bot])

### `golang.org/x/net` updated to v0.62.0

> * fix(deps): update module golang.org/x/net to v0.62.0 [security] (v1.20) (cilium/cilium#49689, @cilium-renovate[bot])

### Security updates bumped to v1.47.0

> * fix(deps): update security updates to v1.47.0 [security] (v1.20) (cilium/cilium#50018, @cilium-renovate[bot])

11 more, none mentioning security

- `* chore(deps): update all-dependencies (v1.20) (cilium/cilium#49898, @cilium-renovate[bot])`
- `* chore(deps): update all-dependencies (v1.20) (cilium/cilium#50031, @cilium-renovate[bot])`
- `* chore(deps): update auto-merge-trusted-deps (v1.20) (cilium/cilium#49598, @cilium-renovate[bot])`
- `* chore(deps): update auto-merge-trusted-deps (v1.20) (cilium/cilium#49748, @cilium-renovate[bot])`
- `* chore(deps): update auto-merge-trusted-deps (v1.20) (cilium/cilium#49899, @cilium-renovate[bot])`
- `* chore(deps): update auto-merge-trusted-deps (v1.20) (cilium/cilium#50024, @cilium-renovate[bot])`
- `* chore(deps): update dependency cilium/cilium-cli to v0.20.2 (v1.20) (cilium/cilium#49896, @cilium-renovate[bot])`
- `* chore(deps): update dependency cilium/cilium-cli to v0.20.3 (v1.20) (cilium/cilium#49997, @cilium-renovate[bot])`
- `* fix(deps): update k8s.io patch updates stable (v1.20) (patch) (cilium/cilium#49897, @cilium-renovate[bot])`
- `* fix(deps): update k8s.io/utils digest to ca04ba5 (v1.20) (cilium/cilium#49754, @cilium-renovate[bot])`
- `* proxy: Bump envoy version to 1.39.x (cilium/cilium#49492, @sayboras)`

## Other changes (26)

- Local IP neighbor entries skipped
- Stable ordering for equal-precedence `GRPCRoute` rules
- BPF load-balancer reconciler backend cleanup fixed
- BPF load-balancer wildcard cleanup uses recorded ownership
- Maglev table cleanup after algorithm switches and restarts
- `CiliumCIDRGroup` update resource version handling fixed
- `cilium/little-vm-helper` dependency updated to v0.0.33
- `sphinxcontrib-googleanalytics` documentation dependency updated to v0.6
- `github/codeql-action` updated to v4.39.2
- `quay.io/cilium/image-tester` Docker tag updated to `v1793531520`
- Stable `lvh-images` dependency patch update
- Docs-builder image builds for qualifying Renovate branches
- Agent warning for unmanaged route MTU with CNI chaining
- WireGuard port reservation guidance added
- Documentation on Standalone DNS Proxy localhost gRPC security
- IPv6 cluster-pool mask-size guidance
- Documentation Makefile no longer builds the `docs-builder` image for command-reference updates
- Helm version flags added to documentation examples and chart-rendering directive
- Tar archive ownership normalized in `render-docs` builds
- Documentation requirements update target fixed
- ztunnel traffic encryption documentation updated
- CopyFromTemplate creates per-copy endpoint state
- SDS abbreviation added to TLS visibility documentation
- Route-MTU warning skipped for `portmap` CNI chaining
- Policy mapstate entry equality uses `derivedFromRules` and `LoggingID()`
- Helm chart image defaults now use digests

---

Summaries are produced by AI, so they can be wrong, and they are provided without warranty.
