Prompt injection, malicious emails...: the wild west of OpenClaw agent system security

An arXiv paper analyzes the security risks of the OpenClaw agent system based on Reddit discussions and GitHub issues, and summarizes seven categories of mitigation strategies. The study analyzed more than 25K posts from 13 OpenClaw-related subreddits and filtered out 162 security-related posts and 1,787 comments; it also collected 358 GitHub issues with security labels and retained 180 issues that describe specific security problems. Through qualitative analysis and evidence-based CWE mapping, the authors identified 521 CWE instances and 59 weakness types, concentrated in broken access control, sensitive data exposure, injection, and unsafe execution. The authors organized 543 practitioner mitigation recommendations into seven categories of strategies, and reconstructed a sensitive data exposure path using fictitious key values, arguing that these problems reflect recurring trust boundary failures in agent execution flows.