Reproduce “BadRobot: Jailbreaking Embodied LLM Agents in the Physical World” 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

Hangtao Zhang, Chenyu Zhu, Xianlong Wang, Ziqi Zhou, Changgan Yin, Minghui Li, Lulu Xue, Yichen Wang and 4 more

Researchers from Huazhong University of Science and Technology, Beihang University, and Griffith University developed BADROBOT, an attack paradigm to 'jailbreak' embodied Large Language Models (LLMs) into performing harmful physical actions. This method significantly increased malicious action success rates by over 200% in digital environments and achieved a 68.57% success rate with real-world robotic manipulators.

Key insights

  • Embodied LLMs are susceptible to 'jailbreak' attacks that induce malicious physical actions, revealing critical safety risks that extend beyond linguistic outputs into the tangible physical world.
  • The research identified three distinct risk surfaces—cascading vulnerability propagation, cross-domain safety misalignment, and conceptual deception—providing a framework for analyzing security issues unique to embodied systems.
  • Current state-of-the-art embodied LLM frameworks, including Voxposer and Code as Policies, are vulnerable to these attacks, indicating they are not yet secure for widespread real-world deployment.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.