OpenClaw HITL Defense Layer: Raise Sandbox Escape Defense from 17% to 91.5% with Human-in-the-Loop Gating

arXiv·medium signal

Shandong University researchers tested OpenClaw across 47 adversarial scenarios and found sandbox escape defenses averaged only 17% (even Claude at 33%). Their proposed HITL defense layer—requiring explicit human approval for high-risk actions—improved overall defense to 91.5% when combined with native protections, intercepting 8 severe attacks that completely bypassed all built-in defenses. The actionable pattern: implement a permission escalation tier in any autonomous agent—classify actions by risk level, auto-approve low-risk, require human confirmation for file system access, network calls, and code execution.