← All updates

Security / News explained

Two EFS mounts could share the wrong local tunnel

AWS’s November 17 efs-utils advisory concerns simultaneous mounts on one host selecting the same local port. The fix spans the mount helper, CSI driver and EKS add-on, with impact bounded by the host’s existing access.

By Zyel · Published · 4 min read

AWS disclosed CVE-2026-130621 on November 17, 2026. The affected components are amazon-efs-utils before 3.3.5, AWS EFS CSI Driver before 3.5.4 and the EKS EFS CSI add-on before 3.5.4-eksbuild.1. The flaw concerns local port selection during simultaneous mounts. [1]

The project advisory describes concurrent mounts on one host choosing the same local port. One mount can then use another mount’s tunnel. An actor able to start mounts on a shared host, such as a tenant able to create pods mounting their EFS volume, may reach another mounted file system under its identity and access point. [2]

AWS bounds the impact to workloads sharing a host at the same time and to access the host already possessed. It says the issue does not disclose the other mount’s IAM credentials or TLS session keys. Cross-account involvement depends on the host owner having configured such a mount. [1]

Our interpretation is that this is a resource-identity failure hidden behind a familiar success signal. A mount can be available and still lead to the wrong storage context. Availability monitoring and a zero exit status answer useful questions, but they do not establish that a workload received its intended data.

The fixed releases were already available

GitHub dates both efs-utils 3.3.5 and CSI Driver 3.5.4 to October 30, 2026. The November 17 bulletin is the later security disclosure. An older dependency record therefore needs its actual version checked rather than being judged by whether it changed on disclosure day. [3] [4]

The advisory names 3.3.5 and 3.5.4 as the respective fixed releases and recommends incorporating the changes into forks or derivatives. The driver bundles the mount helper, which is why the component inventory needs to include the layer actually installed on each host. [2]

The maintainers warn that a crossed mount can report success. Their diagnostic discussion identifies duplicate local ports among NFS mounts and a mount-log pattern of different mounts selecting the same port close together without an intervening unmount. These are specific diagnostic signals, not an assertion that every shared port observation proves an incident. [2]

For a hypothetical acceptance exercise, imagine two authorized test workloads with deliberately distinguishable data. The interesting outcome is whether each workload consistently sees its own intended storage context during overlapping startup. That is a proposed review outcome, not a test we have run against an organisation’s EFS volumes.

Interim limits have an availability cost

AWS says no workaround fully addresses the issue. Widening the mount-port pool and avoiding simultaneous mounts can reduce risk pending an upgrade. The pool must remain below the system’s ephemeral-port range. Serializing node mount operations can increase pod startup time where many volumes need mounting. [1]

That tradeoff belongs in the operational decision. Reduced startup concurrency may be acceptable during a bounded upgrade window, but its effect on recovery or a large deployment should be understood. A temporary risk reduction also needs an owner and a clear endpoint so it does not quietly become the permanent plan.

In our view, the installation record should distinguish the host’s helper package, the node driver and the managed add-on version where used. A control-plane version or a deployment chart name can leave that detail unresolved. The relevant code runs along the mounting path on the host.

The shared-host condition also makes placement relevant to exposure review. The advisory describes a path between workloads that can coexist on a machine and initiate overlapping mounts. An inventory of file systems alone misses the relationship between the workloads, their mount authority and their scheduling.

Check the destination behind the success

This is a useful example of why an infrastructure acceptance result should name the resource it established access to. A log saying mounted, a running pod and a healthy application process can all coexist with a question about data identity. The business effect follows the data a workload actually reads or changes.

A response can therefore connect the version rollout with a review of relevant mount evidence and the intended workload-to-storage mapping. Where evidence is incomplete, that uncertainty should remain visible. The advisory supplies a reason to investigate the named path, not a conclusion that a particular private system suffered unintended access.

The broader engineering implication is concrete: local coordination can be part of an authorization boundary. Two individually legitimate mounts can interact badly when their temporary transport resources collide. Updating the affected components repairs that race; describing acceptance in terms of the intended storage context helps teams verify the outcome they care about.

Sources + further reading

Source dates describe the original material. Publication above is the date this Zyel article went live.

  1. CVE-2026-130621: efs-utils port-selection raceAWS Security · Checked Nov 18, 2026
  2. Concurrent EFS mounts can cross local tunnelsAWS efs-utils maintainers · Checked Nov 18, 2026
  3. Amazon efs-utils 3.3.5 releaseAWS efs-utils maintainers · Checked Nov 18, 2026
  4. AWS EFS CSI Driver 3.5.4 releaseKubernetes SIGs AWS EFS CSI Driver maintainers · Checked Nov 18, 2026

Reader discussion

What would you ask next?

Questions, useful corrections and thoughtful disagreement are welcome. Comments are reviewed before they appear.

No reader comments yet.

10–2,000 characters. Plain text. Leave out passwords, private project data and personal contact details. A blank display name appears as Anonymous.

Complete the CAPTCHA before sending. Verification runs on your device.