Agent Skill Security Exposed

About this event

This live online session examines how a single, seemingly helpful AI skill can become a serious attack vector, enabling code modification, secret exfiltration, memory poisoning, and changes to critical workflows, all while remaining legitimate in the eyes of the developer. Participants will see concrete attack paths using a malicious skill and learn how solutions such as Mitiga’s SkillGate assess skills and configuration files for prompt injection, remote code execution, and data exfiltration techniques, before connecting these findings to runtime defenses across cloud and saas environments. What you will learn • How AI agents change the threat model around skills, hooks, and MCP servers • How a malicious or compromised skill can introduce backdoors, enable secret exfiltration, or poison long‑lived memory • How to use skillgate to scan public repositories and files before an agent loads them • How to combine strict pre‑execution review with runtime detection to contain attacks rather than only record them Who should attend • Security leaders evaluating AI agents and AI-powered development tools • Engineering teams deploying agents into production environments • Security architects designing controls for agent and application security • Professionals responsible for approving or integrating third‑party AI skills and instruction files into production systems Speakers • Idan Cohen - Security Researcher - Mitiga Labs (https://www.linkedin.com/in/idan-cohen-5911aa206/) • Elli Shlomo - Security Researcher - Guardz (https://www.linkedin.com/in/elishlomo/) Notes • Level: practical, technical (200–300) • Language: English • Format: Online, Hebrew, and a formal session with Q&A Connect to the 404 Community • SEC.AI (https://chat.whatsapp.com/LMz1oD1aXsh7JCzJdAiOjm) Group on WhatsApp • Community Notifications (https://chat.whatsapp.com/EFyFFU5zvgiH93XqPH8Q8E) Group on WhatsApp • A landing page for all 404 Community Groups (https://eshlomo.blog/community/)