Aditya Joshi

AI agents are getting alarmingly good at writing code - and then running it. But where does that code actually execute? Far too often it's eval() in your app process or a shared container with a mounted service account token. That's a security incident waiting for a stack trace. This talk introduces Agent Sandbox, a Kubernetes SIG project that turns the cluster into a purpose-built execution layer for untrusted agent code. We'll cover the new CRDs - SandboxTemplate, SandboxWarmPool, and the sandbox router - and show how gVisor isolation, dropped capabilities, and disabled token automounting give you defense in depth. Then we'll go live: spinning up a hardened Python sandbox from a warm pool in under a second, running agent-generated code, and tearing it down. Come for the demo, leave with a threat model.