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

A concerning experiment by security researchers has exposed critical vulnerabilities in AI agent skill verification systems, demonstrating how malicious code could bypass security scanners and potentially compromise thousands of corporate and personal accounts. The controlled study raises significant questions about the current state of AI security and the potential risks organizations face as they increasingly integrate AI agents into their operations.

Security firm AIR created a seemingly legitimate AI agent skill and successfully distributed it through a popular skill marketplace, additionally promoting it via an Instagram advertisement. The researchers reported their creation reached approximately 26,000 agents, including installations on corporate accounts. Perhaps most alarmingly, every security scanner designed to detect malicious skills flagged their creation as safe, allowing it to pass undetected through standard verification processes. The researchers intentionally designed their skill with a harmless payload that merely collected user email addresses to demonstrate the vulnerability without causing actual harm. This controlled experiment was conducted specifically to illustrate how easily a malicious actor could distribute harmful code through trusted channels while evading security measures.

For security teams, these findings carry significant implications. The apparent inability of current security scanners to detect potentially malicious AI agent skills creates a dangerous blind spot in organizational defense strategies. As enterprises increasingly deploy AI agents to handle sensitive business functions, the risk of compromise through seemingly legitimate third-party skills grows substantially. Security professionals must now consider implementing additional verification layers and adopt a more skeptical approach to third-party AI integrations. The experiment also