Tool Specifications Matter: Uncovering and Mitigating Safety Risks in AI Agents