Skill based agents vulnerable to harmful hidden cross skill attacks
Stealth Apart, Harm Together: Skill Cascading Attacks on Skill-Based Agent Systems
Artificial Intelligence
Summary
Skill-based agents use small sets of instructions called skills to perform different tasks. The authors found a new kind of attack where harmful actions are split across several skills, appearing safe individually but dangerous together. They showed that these hidden combined attacks can trick current safety checks and cause serious problems, like hiding drug interaction warnings in medical systems. Their work points out that checking each skill alone is not enough to keep these systems safe.
What this means in practice
- •For software security teams: Detect and prevent multi skill coordinated attacks in modular AI systems using cascaded red teaming benchmarks.
- •For health it developers: Improve safety checks in clinical decision support agents by monitoring combined skill outputs for hidden harmful effects.
Authors
Zihao Zhu, Siwei Lyu, Adel Bibi, Baoyuan Wu
Abstract
A skill is a modular package of natural-language instructions, executable scripts, and reference resources that an agent can load at runtime to extend its capabilities for a specific task. Skill-based agent systems therefore enable flexible reuse of third-party capabilities, but the openness of this skill ecosystem also opens up a new attack surface. Prior work has focused on vulnerabilities within individual skills, but little attention has been paid to risks that arise from interactions across skills. In this paper, we introduce skill cascading attacks, a threat paradigm in which a malicious objective is distributed across multiple skills so that each modification looks benign in isolation, yet their combined execution is harmful. For instance, in a prescription-review pipeline, the first skill weakens signals of recently discontinued medications in the extracted history, the second downgrades the severity of any drug interaction tied to them, and the third suppresses the resulting low-priority alert in the final summary, so that a severe drug-interaction warning silently disappears before reaching the physician. To systematically study this safety blind spot, we develop SkillCascade, an automated multi-agent red-teaming framework, and release SkillCascade-Bench, a benchmark of 213 validated cascading test cases across multiple agent systems and domains. Across representative agents (e.g., OpenClaw, Claude Code, Codex) and LLM backbones, cascaded interactions reliably induce harmful behaviors while evading existing per-skill scanners and runtime monitors. Our findings highlight a gap between component-level integrity and system-level safety, and call for defenses that reason over cross-skill interactions rather than individual skills in isolation.