Improper Authorization in Data Science Pipelines (CVE-2026-18620)

CVE-2026-18620 describes an improper authorization flaw within the Data Science Pipelines component, a key feature of the Kubeflow platform. The vulnerability exists in the setDefaultServiceAccount function, which fails to correctly validate user-provided input during the pipeline execution process. A malicious or compromised tenant with restricted access can deliberately supply a more privileged ServiceAccount (SA) than they are assigned during a CreateRun API request.

If the application fails to restrict this input, the pipeline execution environment is provisioned with the credentials associated with the specified SA rather than the user's default, limited account. This bypass allows an attacker to achieve privilege escalation within the Kubernetes cluster, potentially accessing sensitive secrets or performing unauthorized operations within other users' pods. Defenders should review logs for API requests containing ServiceAccount specifications that deviate from standard tenant provisioning patterns.

Impact

The vulnerability results in unauthorized privilege escalation within the Data Science Pipelines environment. Successful exploitation grants a restricted user access to sensitive information, such as cluster secrets, and enables command execution within the pods of other tenants. This constitutes a severe violation of multi-tenancy isolation, putting the confidentiality and integrity of the entire data science environment at risk.

Recommendation

• Audit Kubernetes RBAC policies and monitor Data Science Pipelines API calls for unexpected ServiceAccount assignment.
• Review all CreateRun API logs to identify requests where the ServiceAccount parameter does not match the requester's expected namespace or project-level role.
• Apply patches provided by the vendor for Data Science Pipelines as soon as they become available.
• Implement Admission Controllers in the Kubernetes cluster to enforce service account restrictions and validate that incoming requests from Data Science Pipelines conform to expected policy constraints.