Fake AI Agent Skill Passed Security Scans and Reportedly Reached 26,000 Agents

This article details an experiment conducted by AIR Security to demonstrate the vulnerabilities within current skill-scanning methods for AI agent skills within popular marketplaces. The firm created a seemingly harmless skill, ‘brand-landingpage,’ that successfully bypassed security scans by leveraging trust signals like GitHub stars and a clean scanner verdict, ultimately reaching 26,000 agents, including some corporate accounts. The core issue highlighted is that scanners only analyze the submitted skill package, failing to detect malicious payloads hosted externally, creating a significant risk for users and organizations relying on these skills.