SCR-Bench (Skill Composition Risk Benchmark) is a benchmark for evaluating security risks that emerge when individually benign agent skills are composed into multi-step workflows. In isolation, each skill appears safe, but harmful outcomes can arise along activated composition paths through capability flow, trust transfer, or authorization… See the full description on the dataset page: https://huggingface.co/datasets/kyle-X1e/SCR-Bench.
