SkillSafetyBench tests agents across 155 artifact attacks

AI45Lab released SkillSafetyBench, a runnable evaluation with 155 prepared cases across six risk domains and 30 categories for attacks hidden in skills, helper scripts, sidecar files, corpora, and other workspace artifacts. The benchmark repository separates ordinary task success from attack_success, while the release report says its public results are project-published rather than independent production evidence. Teams extending portable agent skills should add untrusted local context to eval scope before granting write or execution permissions.