The Research Background section of README.md lists the dataset as 42,447 skills, directly above the 26.1% and 5.2% figures. In the paper those percentages sit on a different denominator.
Liu et al., arXiv:2601.10338 (submitted 15 January 2026), collected 42,447 skills from skills.rest and skillsmp.com, then analysed 31,132 after filtering and deduplication. Section 4.3.1: "Of the 31,132 skills analyzed using SkillScan automated detection, 26.1% contain at least one vulnerability." The detector was the paper's own SkillScan framework.
This travels. AST08 in the OWASP Agentic Skills Top 10 cites the pair as "Per the SkillSpector project, roughly 26.1% of scanned skills contained vulnerabilities", which reads as SkillSpector's own scan results. The section is unchanged since the initial release (7ced4fb, 2026-05-11) and does not link the paper, so nobody downstream can trace it.
Proposed text:
Research Background
Based on "Agent Skills in the Wild: An Empirical Study of Security Vulnerabilities at Scale" (Liu et al., 2026, arXiv:2601.10338):
- Dataset: 42,447 skills collected from two marketplaces, 31,132 analysed after filtering and deduplication
- Vulnerable: 26.1% of the 31,132 contain at least one vulnerability, detected with the paper's SkillScan framework
- High-severity: 5.2% show patterns strongly suggesting malicious intent
- Key finding: Skills with executable scripts are 2.12x more likely to be vulnerable
Line 12 ("Research shows that 26.1% of skills contain vulnerabilities") reads correctly once it links the same paper.
I can send this as a PR with DCO sign-off if that is easier to review.
The Research Background section of README.md lists the dataset as 42,447 skills, directly above the 26.1% and 5.2% figures. In the paper those percentages sit on a different denominator.
Liu et al., arXiv:2601.10338 (submitted 15 January 2026), collected 42,447 skills from skills.rest and skillsmp.com, then analysed 31,132 after filtering and deduplication. Section 4.3.1: "Of the 31,132 skills analyzed using SkillScan automated detection, 26.1% contain at least one vulnerability." The detector was the paper's own SkillScan framework.
This travels. AST08 in the OWASP Agentic Skills Top 10 cites the pair as "Per the SkillSpector project, roughly 26.1% of scanned skills contained vulnerabilities", which reads as SkillSpector's own scan results. The section is unchanged since the initial release (7ced4fb, 2026-05-11) and does not link the paper, so nobody downstream can trace it.
Proposed text:
Research Background
Based on "Agent Skills in the Wild: An Empirical Study of Security Vulnerabilities at Scale" (Liu et al., 2026, arXiv:2601.10338):
Line 12 ("Research shows that 26.1% of skills contain vulnerabilities") reads correctly once it links the same paper.
I can send this as a PR with DCO sign-off if that is easier to review.