The data-labeling interface as a security boundary: task-level privacy and security risks in human-in-the-loop ML pipelines