Security Survey of OpenClaw LLM Agents
A recent survey investigates the security environment surrounding OpenClaw agents, which are open-source frameworks designed for LLM-driven autonomous agents. These agents possess features such as persistent memory, multi-channel communication, and significant autonomy, allowing them to perform intricate multi-step tasks while simultaneously increasing their vulnerability to attacks. Identified threats encompass skill poisoning, cognitive manipulation, cascading failures among multiple agents, and supply-chain weaknesses. This research delves into the unique architecture and attributes that set OpenClaw agents apart from other frameworks.
Key facts
- OpenClaw agents are open-source frameworks for LLM-driven autonomous agents.
- They operate as continuously running systems with persistent memory and multi-channel interaction.
- The attack surface is enlarged due to high-privilege operations and persistent memory.
- Emerging threats include skill poisoning, cognitive manipulation, multi-agent cascading failures, and supply-chain vulnerabilities.
- The survey presents a comprehensive study of the security landscape of OpenClaw agents.
- The paper is available on arXiv with ID 2605.25435.
- The announcement type is new.
- The survey examines the general architecture and key characteristics of OpenClaw agents.
Entities
Institutions
- arXiv