Running untrusted agents on GKE: sandboxing, cold starts, and state

rseroter · x · 2026-11-19

Traditional unit-tested, scanned code in Kubernetes is fine — but agents write and execute code on the fly, breaking the trust model. A diagram breaks down three risks and fixes: (1) unreviewed LLM-generated code on a shared host kernel risks container escape and credential theft → GKE Sandbox (gVisor) or MicroVM (Kata+KVM); (2) multi-second pod spin-up breaks interactive tool loops → pre-warmed SandboxWarmPool with <1s SandboxClaim; (3) idle agents waste RAM but deletion wipes state → Agent Substrate RAM snapshot suspend/restore in <500ms. Includes a live demo of untrusted agents on Agent Substrate with GKE.

Original post →

More from coding & agent

coding & agent channel →