Prevalence of security vulnerabilities in real-world agent skills

Determine the prevalence of security vulnerabilities in real-world agent skills, specifically the modular packages with SKILL.md instructions and optional bundled scripts distributed via public marketplaces such as skills.rest and skillsmp.com, in order to quantify how common these vulnerabilities are across the ecosystem.

Background

The paper highlights that the agent skills ecosystem is new and rapidly expanding, with limited vetting and substantial security risks due to executable scripts and high-trust execution. Prior work focused on LLM prompt-level attacks rather than code and instruction-level vulnerabilities inherent to skills.

The authors note that the community lacks empirical data on the frequency and distribution of vulnerabilities in real-world skills, motivating a large-scale measurement study. This open question is explicitly framed at the outset to justify the research questions and methodology presented later in the paper.