When Interpretability Becomes a Liability: Adversarial Attacks on CBM Concept Layers | ArxivCSExplorer