We let OpenClaw loose on an internal network. Here’s what it found

This report describes a controlled Sophos Red Team experiment that equipped an agentic AI (OpenClaw) with custom, in-house skills and strict ingress/egress and human-in-the-loop guardrails to run noisy penetration testing against a legacy on‑prem network; the test produced 23 actionable findings, significant efficiency gains (e.g., AD reconnaissance reduced from days to hours), demonstrated the value of safety controls, and highlighted friction from model guardrails and broader challenges in securing agentic AI.