Tag
This paper proposes Clinical Guardrail Probes (CG-Probes) to measure risks in patient-facing AI assistants for oncology by recovering guardrail directions from query embeddings using linear probing methods.