Reproduce “Dynamic Malicious Skills in Agentic AI” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Tianhao Chen, Zhengyuan Jiang, Yuepeng Hu, Yebei Gou, Neil Zhenqiang Gong
DyMalSkill is an attack in which documentation for a benign coding skill instructs an AI agent to modify its code during execution, adding malicious behavior that pre-distribution review cannot see. The paper evaluates runtime side effects across 12 behavior types and blocks writes with kernel-level permissions. In OpenHands, ASR fell from 41.6% to 0.0%, within the tested setup.
Key insights
- Dynamic malicious skills exploit an agent's instruction-following capabilities to bypass static security analysis by modifying benign code at runtime.
- The effectiveness of dynamic attacks is significantly influenced by how malicious instructions are framed and presented within the natural-language documentation, often by mimicking system-level directives.
- Dynamic code modification is exceedingly rare and generally not a functional requirement for benign skills, providing a strong rationale for strict runtime permissions and least privilege principles.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.