Building a Digital Supermax: Architecting Air-Gapped Runtimes for Untrusted Agents

Learn how to build secure, air-gapped runtimes for AI agents like OpenClaw. This talk explores network infrastructure and VM design to contain untrusted code, addressing crucial agent security risks.

Overview

I will walk through my design of a system capable of hosting isolated, zero egress, OpenClaw/ClawdBot/MoltBot instances. The system is, in principle, scalable to all AI agents. I can discuss the cloud infrastructure that makes this possible at the network level and still allows OpenClaw to run as root within the VM while allowing end-users to run untrusted code. Think: Codex Cloud but for OpenClaw.

Tech stack