The Authorization Module That Did Not Authenticate: Six Critical Flaws in Dell’s Container Storage Modules
Dell has patched six critical vulnerabilities in Container Storage Modules, two of them rated 10.0. They sit in the component whose job is to enforce authorization, and one of them is a signing secret that appeared in the product documentation.

What Dell published
Dell advisory DSA-2026-448, revision 1.0, dated 1 October 2026, covers multiple vulnerabilities in Container Storage Modules, the software connecting Kubernetes clusters to Dell storage arrays [1]. Six are rated critical, two at the maximum CVSS score of 10.0. No workarounds are documented; the only remediation is version 1.18.0 or later [1][3].
The set is unusual less for the scores than for where the flaws sit: four of the six are in the authorization subsystem or the operator, the components whose function is to decide who may do what to the storage and the cluster.
The two tens: a control plane without a front door
CVE-2026-63688 is described by Dell as a “Missing Authentication for Critical Function vulnerability in the csm-authorization-storage gRPC server” through which “an unauthenticated remote attacker could potentially exploit this vulnerability, leading to unauthorized access to storage backend administrator credentials” [1]. Independent coverage puts the consequence plainly: an attacker can “access administrator credentials for all registered storage arrays” [5], a “complete bypass of the csm-authorization security model” spanning all five supported Dell storage families [2].
CVE-2026-63692 is the same weakness class in the “authorization proxy and tenant service”, where “an unauthenticated network attacker could potentially exploit this vulnerability, leading to bypass of authentication controls” and gain administrative control across all tenants [1][5].
MITRE’s description of the weakness is worth putting next to that. CWE-306 covers a product that “does not perform any authentication for functionality that requires a provable user identity”, and notes that “exposing critical functionality essentially provides an attacker with the privilege level of that functionality” [6]. When the functionality in question is the authorization service for a storage estate, the privilege level of that functionality is the estate.
There is a design point that generalises past Dell. Internal gRPC endpoints between components of one product are routinely built on the assumption that only sibling components reach them. In a cluster that holds only as long as network policy enforces it, and most clusters allow broad pod-to-pod traffic by default.
Secrets in the manual
Two of the remaining four are hard-coded secrets, and the second is the detail worth remembering. CVE-2026-54472 is hard-coded credentials in the CSM Authorization module that let an attacker “forge cryptographically valid administrative JSON Web Tokens” [5]. CVE-2026-61421 is a hard-coded cryptographic key in the JWT authentication component of karavi-authorization, the archived predecessor [1][5].
Older documentation for that component contained the JWT signing secret “supersecret”, and organisations still running it without having changed the documented value “may remain exposed to forged-token attacks and administrative takeover” [5]. Dell’s position on that component is that it “will not be fixed and should be removed” [3].
The weakness classes describe why this is worse than an ordinary credential leak. CWE-798 notes that a hard-coded password “is the same for each installation of the product, and it usually cannot be changed or disabled by system administrators without manually modifying the program” [7]. CWE-321 is blunter still: “If hard-coded cryptographic keys are used, it is almost certain that malicious users will gain access through the account in question” [8].
And a signing key is not a password. RFC 7519 makes the dependency explicit: “the key(s) used to sign and/or encrypt the JWT will typically need to verifiably be under the control of the party identified as the issuer”, and “the contents of a JWT cannot be relied upon in a trust decision unless its contents have been cryptographically secured and bound to the context necessary for the trust decision” [9]. A key printed in a manual is under nobody’s control. Every token the system has ever accepted becomes forgeable, and the verification step still returns success, because the signature is genuinely valid.
The operator is a path to the nodes
CVE-2026-67269, rated 9.9, is improper privilege management in the ContainerStorageModule custom resource reconciler. A low-privileged user can “gain root access to Kubernetes nodes via a malicious ContainerStorageModule resource” [5], or as another account has it, “compromise all nodes in a Kubernetes cluster through a single custom resource submission” [2].
This is the structural risk in the operator pattern, which Kubernetes documents as “software extensions to Kubernetes that make use of custom resources to manage applications and their components”, following the control loop [10]. An operator reads a declarative object submitted by a user and acts on the cluster with its own, much higher privileges. The custom resource is therefore an input to a privileged process, and permission to create one is permission to ask that process to do something.
CVE-2026-67273, at 9.6, is the companion: improper neutralisation of special elements in a template engine, granting “cluster-wide read access to Kubernetes Secrets and the ability to create cluster-scoped RBAC resources” [2][5]. Reading every Secret in the cluster and minting cluster-scoped role bindings are, between them, a durable foothold that outlives the storage software entirely.
Why upgrading is not remediation
Dell’s own instruction settles this. Alongside the upgrade, customers are told to “immediately rotate any JWT signing secrets” [1], and one advisory adds storage credentials to that list [3]. The reason is in the flaws: CVE-2026-63688 exposed storage administrator credentials, and the hard-coded-key issues mean any token minted before the fix may have been forged rather than issued.
So the post-upgrade work has three parts. Rotate the signing secrets, so old tokens stop verifying. Rotate the array credentials the gRPC endpoint could hand out. Then look backwards: review CSM Authorization access logs, audit administrative token usage, and check RBAC policies and custom resources for objects nobody remembers creating [5]. A cluster-scoped role binding created through CVE-2026-67273 is not removed by installing 1.18.0.
What “no known exploitation” is worth
Two advisories state that there is no known exploitation at the time of writing, and that the flaws are “not yet flagged as actively exploited” [3][4]. That is accurate and it is also the weakest kind of reassurance for software of this shape. CSM runs inside clusters, on internal networks, where the telemetry that produces exploitation reports is thinnest. An unauthenticated gRPC call from one pod to another leaves no record at an internet-facing boundary because it never crosses one.
One more divergence worth noting rather than resolving: Dell’s advisory and two accounts give the affected range as versions before 1.17.0, while others state before 1.18.0 [1][2][4][5]. Since the fix is 1.18.0 in every account, the practical answer is unaffected, but anyone recording a version boundary in a compliance note should take it from the advisory.
What to do
- Upgrade Container Storage Modules to 1.18.0 or later. There is no workaround for any of the six [1][3].
- Rotate JWT signing secrets and the storage array credentials that CSM Authorization held, and treat tokens issued before the upgrade as untrustworthy [1][3][9].
- Find and remove any remaining karavi-authorization deployment; it will not be fixed, and check whether the documented signing secret was ever changed [3][5].
- Audit cluster-scoped RBAC objects and ContainerStorageModule custom resources against a known-good list, since a foothold created through these flaws survives the upgrade [2][5][10].
- Restrict who may create or modify the operator’s custom resources, and apply network policy to the authorization service’s internal interfaces rather than relying on cluster-internal reachability [6][10].
- Do not read “no known exploitation” as low risk for software that is only reachable from inside the cluster, where exploitation would be least visible [3][4].
The pattern to take away is narrow and repeatable. When a vulnerability lands in the component that enforces authorization, the blast radius is not the component but everything it was protecting, and the usual recovery step of installing a patch leaves forged credentials and inherited privilege behind. The question to ask of any security control in the estate is not only whether it is patched, but whether it authenticates the callers it was built to trust.
Sources
Every R3KONX article cites its primary material. 10 sources, in order of first citation. Links open the original publication.
- DSA-2026-448: Security Update for Dell Container Storage Modules Multiple Vulnerabilities Dell · 2026-10-01
- Dell CSM Flaws Enable Unauthenticated Admin Access and Root on Kubernetes Nodes The Hacker News · 2026-10-03
- Security advisory: Critical vulnerabilities in Dell Container Storage Modules (CSM) Information Security, University of Toronto · 2026-10-05
- Dell patches critical vulnerabilities in container storage modules SC Media · 2026-10-02
- Critical Dell Container Storage Flaws Let Unauthenticated Attackers Gain Full Administrative Control Cyber Security News · 2026-10-03
- CWE-306: Missing Authentication for Critical Function MITRE · 2026-10-06
- CWE-798: Use of Hard-coded Credentials MITRE · 2026-10-06
- CWE-321: Use of Hard-coded Cryptographic Key MITRE · 2026-10-06
- RFC 7519: JSON Web Token (JWT) IETF · 2015-05-01
- Operator pattern Kubernetes Documentation · 2026-10-06
Researched and written by the R3KONX analysis desk from the cited primary material. CVE identifiers, scores, component names and remediation are taken from Dell's advisory and verified against three independent accounts of it; where those accounts differ on the affected-version boundary, both readings are published. No exploitation detail is reproduced. No exploitation was known at the time of writing, which is a statement about visibility rather than about risk. Corrections to event@r3konx.asia.
