Credential Leakage in LLM Agent Skills: First Large-Scale Empirical Study of 17K Skills

arXiv·high signal

First large-scale empirical study of credential leakage in LLM agent skill ecosystems, analyzing 17,022 skills sampled from 170,226 on SkillsMP using static analysis, sandbox testing, and manual inspection. Identified 520 vulnerable skills with 1,708 issues across a taxonomy of 10 leakage patterns (4 accidental, 6 adversarial). Third-party skills routinely handle sensitive credentials in privileged environments with poorly understood leakage surfaces.