Skills are modular packages of instructions, scripts, and resources that agents load at runtime, and they all share one context window. That means content one skill writes can be read as input by the next skill, so risk can propagate across skills that are independently vetted. This work introduces skill cascading attacks, where an attacker splits a malicious objective across two or more skills so each modified skill looks benign on its own, but their combined execution is harmful.
The authors built SkillCascade, a multi-agent red-teaming framework that generates and validates such attacks, and released SkillCascade-Bench with 213 validated test cases across 10 domains. Across three agent systems and eight LLM backbones, cascaded interactions reliably induced harmful behavior while evading existing per-skill scanners and runtime monitors. This matters because current defenses focus on individual skills, while the actual risk lives in cross-skill interaction.