AI RESEARCH
ClawTrap: A MITM-Based Red-Teaming Framework for Real-World OpenClaw Security Evaluation
arXiv CS.AI
•
ArXi:2603.18762v1 Announce Type: cross Autonomous web agents such as \textbf{OpenClaw} are rapidly moving into high-impact real-world workflows, but their security robustness under live network threats remains insufficiently evaluated. Existing benchmarks mainly focus on static sandbox settings and content-level prompt attacks, which leaves a practical gap for network-layer security testing. In this paper, we present \textbf{ClawTrap}, a \textbf{MITM-based red-teaming framework for real-world OpenClaw security evaluation.