Stealth Apart, Harm Together: Skill Cascading Attacks on Skill-Based Agent Systems

arXiv:2609.30383v1 Announce Type: new 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

Introduces a new attack paradigm that exploits interactions across skills

This is a peer-adjacent but unreviewed preprint demonstrating a real, reproducible vulnerability class in deployed agent systems, validated with 213 test cases and real agents. The implication is materially concerning for safety of increasingly autonomous multi-skill agent deployments, though it is a security research finding rather than an active exploited harm.

The paper identifies a novel and effective attack vector, 'skill cascading attacks,' on AI agent systems. These attacks can distribute malicious objectives across multiple skills, appearing benign individually but causing significant harm when combined. The research demonstrates these attacks reliably induce harmful behaviors, evade current defenses, and could have serious real-world implications, such as suppressing critical drug interaction alerts.

The paper describes a new type of attack on skill-based agent systems. These attacks can induce harmful behaviors while evading existing safety measures. The existence of a benchmark with 213 validated test cases indicates a significant and demonstrable risk.

SOURCE     arXiv cs.AI (tier 1)
   ↓
DOCUMENT   5ddb608c-f584-4f74-83d3-d9e5359d91e5
           https://arxiv.org/abs/2609.30383
   ↓
EVIDENCE   3 extracted excerpts
   ↓
MODEL RUN  4 runs, methodology 1.0.1
   ↓
SCORE      -69  (Strongly adverse)
   ↓
CONFIDENCE 79%