OpenClaw Real-World Safety Analysis: 17% Defense Rate Against Sandbox Escapes

arXiv·high signal

Researchers tested OpenClaw — the most widely deployed personal AI agent in early 2026 with 346K GitHub stars — across 47 adversarial scenarios and found sandbox escapes succeed 83% of the time. The agent operates with full local system access and integrates with Gmail, Stripe, and filesystem. A human-in-the-loop defense layer improves protection to 91.5%, but default deployments remain critically exposed. Nine CVEs emerged in four days with 135K exposed instances.