A vulnerability in the Trainer v2 TrainJob API allows users with permissions to create TrainJobs to inject valueFrom.secretKeyRef into environment variable definitions. This can lead to exfiltration of sensitive secrets from the namespace and privilege escalation if the TrainJob RBAC is broader than the Secret read RBAC. The issue affects a Technology Preview feature in Red Hat OpenShift AI and has a low impact. Mitigation involves aligning TrainJob creation permissions with Secret read permissions or restricting the valueFrom.secretKeyRef field via a validating webhook.