Three security scanners disagree on 23,702 of the 61,990 agent skills they all cover, and 85% of readable skills carry privilege evidence

arXiv·high signalXBlueskyLinkedInCopy link

A measurement study of the OpenClaw skill registry using Git history, GitHub issues/PRs and three ClawHub snapshots found the observable skill stock nearly doubled in 91 days before creation rates fell from their spring peak. Attention is brutally concentrated (top 10% of skills took 46.93% of downloads), human scrutiny is absent (77.86% have zero stars and zero comments) and 85.06% of readable skills carry evidence of shell, network, credential, file or process privilege. Automated cleanup is not a substitute: the three scanners disagreed on 23,702 of 61,990 commonly-covered skills, and after human adjudication their weighted sensitivity against the reference standard ranged from just 21.67% to 61.06% — meaning a single scanner badge on a marketplace skill is close to meaningless.