Trojan’s Whisper: Stealthy Manipulation of OpenClaw through Injected Bootstrapped Guidance

Fazhong Liu, Zhuoyan Chen, Tu Lan, Haozhen Tan, Zhenyu Xu, Xiang Li, Guoxing Chen, Yan Meng, and Haojin Zhu.

This paper studies guidance injection, an attack surface in OpenClaw’s extensible skill ecosystem. Adversarial narratives embedded in bootstrap guidance files frame harmful actions as routine practices and influence subsequent agent behavior.

Using ORE-Bench, the study evaluates 26 malicious skills across 13 attack categories, 52 natural user prompts, and six LLM backends. Reported attack success rates range from 16.0% to 64.2%, while 94% of the malicious skills evade the evaluated static and LLM-based scanners. The findings motivate capability isolation, runtime policy enforcement, and transparent guidance provenance.

Recommended citation: Fazhong Liu, Zhuoyan Chen, Tu Lan, Haozhen Tan, Zhenyu Xu, Xiang Li, Guoxing Chen, Yan Meng, and Haojin Zhu. "Trojan's Whisper: Stealthy Manipulation of OpenClaw through Injected Bootstrapped Guidance." arXiv:2603.19974, 2026. doi: 10.48550/arXiv.2603.19974.
Download Paper