Summary
GPUStack's cluster API proxy could forward an authenticated non-administrator's requests to a Kubernetes cluster without checking authorization for the requested resource. If the deployment conditions are met, the requests may be carried out using a worker's Kubernetes ServiceAccount permissions, enabling access to or changes to cluster resources.
Affected context
The issue affects GPUStack's cluster API proxy when used with a Kubernetes-provider cluster that has a reachable worker. In this context, the proxy used that worker's Kubernetes ServiceAccount credentials to make requests, without first verifying that the authenticated GPUStack user was authorized to perform the requested action on the target resource.
Exploit preconditions
- Attack surface: GPUStack's cluster API proxy for a Kubernetes-provider cluster
- Required access: An authenticated non-administrator account able to reach the proxy
- Environment condition: The selected cluster must have a reachable worker whose ServiceAccount credentials are used by the proxy
- Trigger point: A request forwarded by the proxy to the selected cluster
- Why exploitation is possible: The proxy does not enforce resource-level authorization for the caller, so requests may be performed with the worker ServiceAccount's permissions
Why this happens
The proxy delegated Kubernetes API requests using a worker's ServiceAccount credentials without checking whether the requesting GPUStack user could access or modify the requested resource. This is a missing authorization check, mapping to CWE-862: Missing Authorization. The use of a caller-influenced target or resource may also relate to CWE-639: Authorization Bypass Through User-Controlled Key when authorization is not correctly bound to that selection.
Potential impact
- A caller may read sensitive Kubernetes resources permitted to the ServiceAccount
- A caller could create or modify workloads within the ServiceAccount's permissions
- A caller may delete resources the ServiceAccount is authorized to delete
- Attacker-controlled workloads could enable code execution in the target cluster, depending on workload permissions and cluster controls
General mitigation guidance
Defensive designs typically authorize each request against the authenticated user's identity, target cluster, resource, and requested action before proxying it.
Common security patterns include:
- Applying explicit, resource-level authorization checks to proxied operations
- Binding cluster and resource selections to the caller's authorized scope
- Using narrowly scoped ServiceAccount permissions and least-privilege identities
- Denying requests by default when the caller's authorization cannot be established
- Auditing and limiting proxy operations that can create, modify, or delete resources
How teams detect issues like this
High-level detection approaches include:
- Reviewing proxy code for forwarded requests that rely on service credentials without caller-specific authorization
- Using static analysis and code review checks to flag proxy handlers missing authorization decisions
- Testing access policies across user roles, resource types, actions, and selected clusters
- Enforcing policy checks in CI or API gateways so requests are denied unless explicitly authorized
Audit logs can also help identify non-administrator requests that perform sensitive Kubernetes operations through the proxy.

