Same Payload, Different Channel: Measuring Trust Asymmetry in Tool-Using Language Models | ArxivCSExplorer