newsfilter
← Interviews
AI Engineer15 min

You Can't Tell What an Agent Skill Will Do by Reading It - Roey Ben Chaim, Zenity

Roey Ben Chaim

Conference Presentation, Keynote

The Zenity Labs research team developed a "Detonation Chamber" sandbox to shift agent skill security from static code analysis to runtime execution testing, revealing how malicious payloads mimic legitimate behaviors like InfoStealers through multi-stage chains and social engineering. This approach successfully identified and neutralized 100 harmful skills, including a "Carly Paperclip" impersonation campaign that harvested credentials across 1.7 million downloads before Vercel and Microsoft removed the assets. The findings redefine agent context risk to include external dependencies and necessitate strict runtime monitoring, with the detection methodology now available for public use via Artoto.io.