Security Best Practices

Istio security features provide strong identity, powerful policy, transparent TLS encryption, and authentication, authorization and audit (AAA) tools to protect your services and data. However, to fully make use of these features securely, care must be taken to follow best practices. It is recommended to review the Security overview before proceeding.

Mutual TLS

Istio will automatically encrypt traffic using Mutual TLS whenever possible. However, proxies are configured in permissive mode by default, meaning they will accept both mutual TLS and plaintext traffic.

While this is required for incremental adoption or allowing traffic from clients without an Istio sidecar, it also weakens the security stance. It is recommended to migrate to strict mode when possible, to enforce that mutual TLS is used.

Mutual TLS alone is not always enough to fully secure traffic, however, as it provides only authentication, not authorization. This means that anyone with a valid certificate can still access a service.

To fully lock down traffic, it is recommended to configure authorization policies. These allow creating fine-grained policies to allow or deny traffic. For example, you can allow only requests from the app namespace to access the hello-world service.

Authorization policies

Istio authorization plays a critical part in Istio security. It takes effort to configure the correct authorization policies to best protect your clusters. It is important to understand the implications of these configurations as Istio cannot determine the proper authorization for all users. Please follow this section in its entirety.

Apply default-deny authorization policies

We recommend you define your Istio authorization policies following the default-deny pattern to enhance your cluster’s security posture. The default-deny authorization pattern means your system denies all requests by default, and you define the conditions in which the requests are allowed. In case you miss some conditions, traffic will be unexpectedly denied, instead of traffic being unexpectedly allowed. The latter typically being a security incident while the former may result in a poor user experience, a service outage or will not match your SLO/SLA.

For example, in the authorization for HTTP traffic task, the authorization policy named allow-nothing makes sure all traffic is denied by default. From there, other authorization policies allow traffic based on specific conditions.

Customize your system on path normalization

Istio authorization policies can be based on the URL paths in the HTTP request. Path normalization (a.k.a., URI normalization) modifies and standardizes the incoming requests’ paths, so that the normalized paths can be processed in a standard way. Syntactically different paths may be equivalent after path normalization.

Istio supports the following normalization schemes on the request paths, before evaluating against the authorization policies and routing the requests:

NONENo normalization is done. Anything received by Envoy will be forwarded exactly as-is to any backend service.../%2Fa../b is evaluated by the authorization policies and sent to your service.
BASEThis is currently the option used in the default installation of Istio. This applies the normalize_path option on Envoy proxies, which follows RFC 3986 with extra normalization to convert backslashes to forward slashes./a/../b is normalized to /b. \da is normalized to /da.
MERGE_SLASHESSlashes are merged after the BASE normalization./a//b is normalized to /a/b.
DECODE_AND_MERGE_SLASHESThe most strict setting when you allow all traffic by default. This setting is recommended, with the caveat that you will need to thoroughly test your authorization policies routes. Percent-encoded slash and backslash characters (%2F, %2f, %5C and %5c) are decoded to / or \, before the MERGE_SLASHES normalization./a%2fb is normalized to /a/b.

To emphasize, the normalization algorithms are conducted in the following order:

  1. Percent-decode %2F, %2f, %5C and %5c.
  2. The RFC 3986 and other normalization implemented by the normalize_path option in Envoy.
  3. Merge slashes

Examples of configuration

Ensuring Envoy normalizes request paths to match your backend services’ expectation is critical to the security of your system. The following examples can be used as reference for you to configure your system. The normalized URL paths, or the original URL paths if NONE is selected, will be:

  1. Used to check against the authorization policies
  2. Forwarded to the backend application
Your application…Choose…
Relies on the proxy to do normalizationBASE, MERGE_SLASHES or DECODE_AND_MERGE_SLASHES
Normalizes request paths based on RFC 3986 and does not merge slashesBASE
Normalizes request paths based on RFC 3986, merges slashes but does not decode percent-encoded slashesMERGE_SLASHES
Normalizes request paths based on RFC 3986, decodes percent-encoded slashes and merges slashesDECODE_AND_MERGE_SLASHES
Processes request paths in a way that is incompatible with RFC 3986NONE

How to configure

You can use istioctl to update the mesh config:

$ istioctl upgrade --set meshConfig.pathNormalization.normalization=DECODE_AND_MERGE_SLASHES

or by altering your operator overrides file

$ cat <<EOF > iop.yaml
kind: IstioOperator
      normalization: DECODE_AND_MERGE_SLASHES
$ istioctl install -f iop.yaml

Alternatively, if you want to directly edit the mesh config, you can add the pathNormalization to the mesh config, which is the istio-<REVISION_ID> configmap in the istio-system namespace. For example, if you choose the DECODE_AND_MERGE_SLASHES option, you modify the mesh config as the following:

apiVersion: v1
    mesh: |-
        normalization: DECODE_AND_MERGE_SLASHES

Less common normalization configurations

Case Normalization

In some environments, it may be useful to have paths in authorization policies compared in a case insensitive manner. For example, treating https://myurl/get and https://myurl/GeT as equivalent. In those cases, the EnvoyFilter shown below can be used. This filter will change both the path used for comparison and the path presented to the application.

kind: EnvoyFilter
  name: ingress-case-insensitive
  namespace: istio-system
  - applyTo: HTTP_FILTER
      context: GATEWAY
            name: ""
              name: "envoy.filters.http.router"
      operation: INSERT_BEFORE
        name: envoy.lua
            "@type": ""
            inlineCode: |
              function envoy_on_request(request_handle)
                local path = request_handle:headers():get(":path")
                request_handle:headers():replace(":path", string.lower(path))

Understand traffic capture limitations

The Istio sidecar works by capturing both inbound traffic and outbound traffic and directing them through the sidecar proxy.

However, not all traffic is captured:

  • Redirection only handles TCP based traffic. Any UDP or ICMP packets will not be captured or modified.
  • Inbound capture is disabled on many ports used by the sidecar as well as port 22. This list can be expanded by options like
  • Outbound capture may similarly be reduced through settings like or other means.

In general, there is minimal security boundary between an application and its sidecar proxy. Configuration of the sidecar is allowed on a per-pod basis, and both run in the same network/process namespace. As such, the application may have the ability to remove redirection rules and remove, alter, terminate, or replace the sidecar proxy. This allows a pod to intentionally bypass its sidecar for outbound traffic or intentionally allow inbound traffic to bypass its sidecar.

As a result, it is not secure to rely on all traffic being captured unconditionally by Istio. Instead, the security boundary is that a client may not bypass another pod’s sidecar.

For example, if I run the reviews application on port 9080, I can assume that all traffic from the productpage application will be captured by the sidecar proxy, where Istio authentication and authorization policies may apply.

Defense in depth with NetworkPolicy

To further secure traffic, Istio policies can be layered with Kubernetes Network Policies. This enables a strong defense in depth strategy that can be used to further strengthen the security of your mesh.

For example, you may choose to only allow traffic to port 9080 of our reviews application. In the event of a compromised pod or security vulnerability in the cluster, this may limit or stop an attackers progress.

Securing egress traffic

A common misconception is that options like outboundTrafficPolicy: REGISTRY_ONLY acts as a security policy preventing all access to undeclared services. However, this is not a strong security boundary as mentioned above, and should be considered best-effort.

While this is useful to prevent accidental dependencies, if you want to secure egress traffic, and enforce all outbound traffic goes through a proxy, you should instead rely on an Egress Gateway. When combined with a Network Policy, you can enforce all traffic, or some subset, goes through the egress gateway. This ensures that even if a client accidentally or maliciously bypasses their sidecar, the request will be blocked.

Configure TLS verification in Destination Rule when using TLS origination

Istio offers the ability to originate TLS from the sidecar proxy. This enables applications that send plaintext HTTP traffic to be transparently “upgraded” to HTTPS.

Care must be taken when configuring the DestinationRule’s tls setting to specify the caCertificates field. When this is not set, the servers certificate will not be verified.

For example:

kind: DestinationRule
  name: google-tls
      mode: SIMPLE
      caCertificates: /etc/ssl/certs/ca-certificates.crt


When running an Istio gateway, there are a few resources involved:

  • Gateways, which controls the ports and TLS settings for the gateway.
  • VirtualServices, which control the routing logic. These are associated with Gateways by direct reference in the gateways field and a mutual agreement on the hosts field in the Gateway and VirtualService.

Restrict Gateway creation privileges

It is recommended to restrict creation of Gateway resources to trusted cluster administrators. This can be achieved by Kubernetes RBAC policies or tools like Open Policy Agent.

Avoid overly broad hosts configurations

When possible, avoid overly broad hosts settings in Gateway.

For example, this configuration will allow any VirtualService to bind to the Gateway, potentially exposing unexpected domains:

- port:
    number: 80
    name: http
    protocol: HTTP
  - "*"

This should be locked down to allow only specific domains or specific namespaces:

- port:
    number: 80
    name: http
    protocol: HTTP
  - "" # Allow only VirtualServices that are for
  - "default/" # Allow only VirtualServices in the default namespace that are for
  - "route-namespace/*" # Allow only VirtualServices in the route-namespace namespace for any host

Isolate sensitive services

It may be desired to enforce stricter physical isolation for sensitive services. For example, you may want to run a dedicated gateway instance for a sensitive, while utilizing a single shared gateway instance for less sensitive domains like and

Explicitly disable all the sensitive http host under relaxed SNI host matching

It is reasonable to use multiple Gateways to define mutual TLS and simple TLS on different hosts. For example, use mutual TLS for SNI host and simple TLS for SNI host *

kind: Gateway
  name: guestgateway
    istio: ingressgateway
  - port:
      number: 443
      name: https
      protocol: HTTPS
    - "*"
      mode: SIMPLE
kind: Gateway
  name: admingateway
    istio: ingressgateway
  - port:
      number: 443
      name: https
      protocol: HTTPS
      mode: MUTUAL

If the above is necessary, it’s highly recommended to explicitly disable the http host in the VirtualService that attaches to * The reason is that currently the underlying envoy proxy does not require the http 1 header Host or the http 2 pseudo header :authority following the SNI constraints, an attacker can reuse the guest-SNI TLS connection to access admin VirtualService. The http response code 421 is designed for this Host SNI mismatch and can be used to fulfill the disable.

kind: VirtualService
  name: disable-sensitive
  - ""
  - guestgateway
  - match:
    - uri:
        prefix: /
          value: 100
        httpStatus: 421
    - destination:
          number: 8000
        host: dest.default.cluster.local

Protocol detection

Istio will automatically determine the protocol of traffic it sees. To avoid accidental or intentional miss detection, which may result in unexpected traffic behavior, it is recommended to explicitly declare the protocol where possible.


In order to transparently capture all traffic, Istio relies on iptables rules configured by the istio-init initContainer. This adds a requirement for the NET_ADMIN and NET_RAW capabilities to be available to the pod.

To reduce privileges granted to pods, Istio offers a CNI plugin which removes this requirement.

Use hardened docker images

Istio’s default docker images, including those run by the control plane, gateway, and sidecar proxies, are based on ubuntu. This provides various tools such as bash and curl, which trades off convenience for an increase attack surface.

Istio also offers a smaller image based on distroless images that reduces the dependencies in the image.

Release and security policy

In order to ensure your cluster has the latest security patches for known vulnerabilities, it is important to stay on the latest patch release of Istio and ensure that you are on a supported release that is still receiving security patches.

Avoid alpha and experimental features

All Istio features and APIs are assigned a feature status, defining its stability, deprecation policy, and security policy.

Because alpha and experimental features do not have as strong security guarantees, it is recommended to avoid them whenever possible.

To determine the feature status of features in use in your cluster, consult the Istio features list.

Configure third party service account tokens

To authenticate with the Istio control plane, the Istio proxy will use a Service Account token. Kubernetes supports two forms of these tokens:

  • Third party tokens, which have a scoped audience and expiration.
  • First party tokens, which have no expiration and are mounted into all pods.

Because the properties of the first party token are less secure, Istio will default to using third party tokens. However, this feature is not enabled on all Kubernetes platforms.

If you are using istioctl to install, support will be automatically detected. This can be done manually as well, and configured by passing --set or --set

To determine if your cluster supports third party tokens, look for the TokenRequest API. If this returns no response, then the feature is not supported:

$ kubectl get --raw /api/v1 | jq '.resources[] | select(.name | index("serviceaccounts/token"))'
    "name": "serviceaccounts/token",
    "singularName": "",
    "namespaced": true,
    "group": "",
    "version": "v1",
    "kind": "TokenRequest",
    "verbs": [

While most cloud providers support this feature now, many local development tools and custom installations may not prior to Kubernetes 1.20. To enable this feature, please refer to the Kubernetes documentation.

Was this information useful?
Do you have any suggestions for improvement?

Thanks for your feedback!