Security scanners for AI agent skill marketplaces fail to detect malicious skills, researchers find

Summary

The article exposes critical security flaws in AI agent skill marketplaces, where malicious skills designed to steal credentials, exfiltrate data, and hijack agents are proliferating. The authors tested multiple security scanners—including ClawHub's detector, Cisco's agent skill scanner, and three scanners on skills.sh—and successfully bypassed all of them with minimal effort. The findings reveal that current security measures for AI agent ecosystems are fundamentally inadequate, as the authors conceived and implemented three out of four malicious skills in under an hour.