AI
Study reveals fragility of fast-growing AI agent skill registries despite security screening
A new paper examining the OpenClaw AI agent ecosystem finds that rapid growth in skill registries has outpaced governance, with security scanners disagreeing sharply and few checks preventing risky capabilities from spreading.
AI agents that execute real-world actions—from system commands to credential manipulation—increasingly rely on modular "skills" that are distributed through public registries. A new research paper from arXiv examines what happened to the OpenClaw agent ecosystem during a period of explosive growth, revealing governance gaps and inconsistent security screening that may pose risks as these systems scale.
The Viral Boom and Its Aftermath#
In the first half of 2026, the OpenClaw registry experienced a surge in submissions. The number of available skills roughly doubled in just over three months, with the majority of new entries added within a compressed two-month window. However, by the time researchers completed their analysis, the growth momentum had slowed markedly, with monthly creation rates and repository activity declining from their springtime peaks. The study draws on Git history, GitHub issues and pull requests, and snapshots of the ClawHub registry to assess what remains after the wave subsides.
Concentration and the Long Tail Problem#
Downloads are starkly concentrated: the most popular 10% of skills capture nearly 47% of all usage, leaving the majority of offerings with minimal adoption. Surprisingly, traditional predictors of engagement—such as code size or download volume—failed to forecast whether skills would remain active once the researchers accounted for their creation date and age. This suggests the registry lacks a reliable method for identifying which entries warrant attention or removal.
Human Oversight Largely Absent#
The research documents a troubling pattern regarding human scrutiny. About 78% of skills carry zero star ratings and zero user comments, indicating minimal community review. More concerning, roughly 85% of readable skills contain evidence of privilege escalation or dangerous capabilities—shell access, network operations, credential handling, file manipulation, or process control. Yet these persist with little friction.
Security Scanners Disagree Sharply#
Automated security screening should fill the gap left by absent human curation, but the analysis reveals a significant problem: three independent security scanners disagreed on the classification of nearly 24,000 of the 62,000 skills they all evaluated. After human experts manually reviewed disputed cases, the scanners' sensitivity scores ranged widely—from 22% to 61%—indicating they caught inconsistent fractions of actual threats. This disparity raises doubt about whether relying on any single tool can meaningfully govern the registry.
