Intel / article

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

Source: arxiv.org
Author: Zihao Zhu, Siwei Lyu, Adel Bibi, Baoyuan Wu
Date: Published Sep 28, 2026

Context

Checking an 's reusable skills separately can miss harm created by their interaction. SkillCascade studies attacks that distribute a malicious objective across multiple packages of instructions, scripts and reference material, with each modification appearing benign in isolation.

The paper's prescription-review example illustrates the chain: one step weakens medication-history information, another lowers the severity of a related interaction warning, and a later step omits the downgraded warning. The danger emerges from what each step passes to the next. The authors report a of 213 validated cascading cases and failures of tested individual- scanners and runtime monitors. Their finding supports evaluating the combined workflow and its intermediate information, alongside checks on each component.

Key takeaways · AI-distilled

Skill cascading attacks split a malicious goal across several agent skills so each change looks harmless alone, while their combined execution causes harm.
The paper's example is a prescription-review pipeline where three skills each slightly weaken, downgrade or suppress a signal, so a severe drug-interaction warning never reaches the physician.
The authors release SkillCascade, an automated red-teaming framework, and SkillCascade-Bench with 213 validated cascading cases across several agent systems and domains.
Against agents including OpenClaw, Claude Code and Codex, cascades reliably induced harmful behavior while evading per-skill scanners and runtime monitors, so defenses must reason across skill interactions.

Terms in this piece · Glossary

AI agent — An AI system that doesn't just answer once but works toward a goal in a loop — taking actions, reading the results, and deciding what to do next.
benchmark — A standard public test set for comparing AI models — the shared scoreboards behind every "model X beats model Y" claim.
agent skill — A reusable instruction file that teaches an agent how to do one job well — the procedure, the tools, and what counts as done.
multi-agent — Using several AI agents on one problem — splitting work in parallel, checking each other, or filling different roles like planner and reviewer.

Read the source arxiv.org