VirusTotal: Some Malicious AI Skills Hide Instructions

VirusTotal analyzed 35,878 skills for AI agents and found 6,637 to be malicious. In 62% of them, the malicious payload was written not as code, but as ordinary text in the skill documentation.

The researchers separately examined 90 skills in which instructions were hidden from ordinary view: 56 used long sequences of blank lines, 28 placed text in HTML comments, and six used invisible Unicode characters. In some cases, a file opened in an editor or browser looked different from its contents as seen by the AI agent.

VirusTotal notes that these concealment methods are still rare. As checks, the company recommends reading the entire source file, normalizing text before scanning, and not allowing skills to modify the agent’s memory and configuration files without the user’s explicit permission.

Image: Polk, W. Timothy
Bassham, Lawrence E., III. Public domain

Translation editors: Соловьев В., Пеков А.