What question did this study set out to answer?

The aim is to conduct an extensive security audit of AI agent skills to identify vulnerabilities and malware.

April 17, 20260 citationsOpen Access

96,096 Skills, 751 Malware: A Large-Scale Security Audit of the AI Agent Ecosystem

Key Points

The aim is to conduct an extensive security audit of AI agent skills to identify vulnerabilities and malware.
Scanned 96,096 skills across six registries using 113 ATR detection rules.
Flagged a total of 1,302 skills for review, representing 1.35% of total scanned.
Confirmed 751 instances of malware from coordinated threat actors.
Identified malware making up 0.78% of the total skills scanned.
Tool description poisoning was identified as the cause for 53% of malware detections.
Achieved a median scan latency of 5.39ms per skill.

Abstract

The largest automated security audit of AI agent skills to date. 96,096 skills scanned across six registries (OpenClaw, ClawHub, Skills.sh, Hermes Agent, MCP Registry) using 113 ATR detection rules. 1,302 flagged (1.35%), 751 confirmed malware (0.78%) from three coordinated threat actors. Tool description poisoning accounts for 53% of detections. Median scan latency 5.39ms per skill. All findings reported and blacklisted.

Demander à l'IA

Bookmark

View Full Paper