CubeSandbox

Tracked

A high-performance, hardware-isolated sandbox service for AI agents built on RustVMM and KVM, with E2B SDK compatibility and sub-60ms cold starts.

Author Tencent Cloud Open Sourced 2026-04-10 Last Commit Unknown

Overview

CubeSandbox is an instant, concurrent, secure, and lightweight sandbox service designed specifically for AI agents. Built on RustVMM and KVM, it provides true kernel-level isolation where each agent runs with its own dedicated guest OS kernel — eliminating container escape risks inherent in shared-kernel approaches like Docker.

Key Features

  • Sub-60ms cold start for fully serviceable sandboxes via resource pool pre-provisioning and snapshot cloning
  • Ultra-low memory overhead below 5MB per instance using CoW technology and a Rust-rebuilt stripped runtime
  • Hardware-level kernel isolation with dedicated guest OS per sandbox, plus eBPF-based network security (CubeVS)
  • Drop-in E2B SDK compatibility — swap one URL to migrate from closed-source sandboxes
  • Event-level snapshot, clone, and rollback at hundred-millisecond granularity via CubeCoW engine
  • Single-node and multi-node cluster deployment, validated at Tencent Cloud production scale

Use Cases

  • Secure code execution for AI coding agents and LLM-generated code
  • Browser automation in isolated environments
  • Reinforcement learning training environments (SWE-Bench, etc.)
  • High-density multi-tenant agent hosting on bare-metal infrastructure

Technical Details

  • Built on RustVMM virtualization with KVM hardware acceleration
  • Copy-on-Write (CubeCoW) snapshot engine for instant cloning and state rollback
  • eBPF-powered CubeVS for inter-sandbox network isolation and egress filtering
  • Available as PyPI package (cubesandbox) for Python integration