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

Purpose This paper aims to investigate privacy and security vulnerabilities arising from human-in-the-loop (HITL) interactions during machine learning data labeling and quality assurance/review. It argues that the data-labeling interface constitutes a distinct and underexamined security boundary, separate from infrastructure controls and organizational compliance frameworks. Design/methodology/approach The study adopts a qualitative, practice-oriented research design combining thematic analysis of publicly available documentation and workflows with responses from six practitioners gathered via a structured online instrument. Findings are organized through a task-level vulnerability taxonomy and a design–exposure–control matrix. Findings Six recurring vulnerability families are identified: excess contextual visibility, metadata exposure, privilege expansion in QA/review, instruction-level disclosure, inference-based exposure and weak anti-exfiltration controls. The matrix shows that many existing controls are partial rather than absent, and that governance principles such as least privilege and purpose limitation are frequently undermined at the task-interface level. Research limitations/implications The study relies on publicly available documentation and a small purposive sample of six practitioners, which limits statistical generalizability. Findings may not fully capture proprietary workflows. Future work should compare findings across sectors such as healthcare and financial services and test specific interface interventions empirically. Practical implications The safeguards checklist and design–exposure–control matrix offer practitioners concrete tools to audit and improve task-level privacy and security in labeling environments without requiring changes to underlying infrastructure or compliance frameworks. Originality/value To the best of the authors’ knowledge, this study offers the first systematic treatment of the data-labeling interface as a task-level security boundary. It provides a vulnerability taxonomy, a design–exposure–control matrix and an actionable safeguards checklist, offering practitioners and researchers structured tools to evaluate and strengthen privacy and security in HITL data-preparation environments.