OpenSkillRisk: Benchmarking Agent Safety When Using Real-World Risky Third-Party Skills