Benign in Isolation, Harmful in Composition: Security Risks in Agent Skill Ecosystems
A new research paper introduces SCR-Bench, a benchmark designed to evaluate security risks in LLM agent skill ecosystems. The research highlights that while individual skills may appear safe in isolation, their composition can lead to significant security vulnerabilities, such as data leakage and unauthorized operations. SCR-Bench measures these risks by analyzing downstream state changes and path-level outcomes across composed skill executions, revealing that composed paths can expose risks largely absent in isolated evaluations. AI
IMPACT Highlights the need for path-aware security assessments in LLM agents, potentially influencing future agent development and security practices.