Breaking

Tencent Cloud Open-Sources CubeSandbox for AI Agents Under Apache 2.0

July 6, 2026 at 15:45 EDT

AI Agents
Open Source
Security

Tencent Cloud has open-sourced CubeSandbox, a microVM-based sandbox for running AI agent code securely and quickly, under the Apache 2.0 license. It boots in under 60 milliseconds and consumes under 5MB of memory per instance, and it assigns a dedicated guest kernel to each sandbox for hardware-level isolation.

April 2026 · Tencent Cloud

CubeSandbox goes open source: microVM speed for AI agents

An Apache 2.0 sandbox for running AI agent code that claims sub-60ms cold starts and under 5MB memory per instance — while keeping KVM hardware-level isolation.

<60ms

Avg. single cold start

<5MB

Memory overhead per instance

100K+

Instances launched per minute, platform-wide

Claimed startup speed vs. Docker

Company measurements on bare metal — a range of 2.5× to 50× faster

1×

Docker
(baseline)

2.5×

CubeSandbox
(low end)

50×

CubeSandbox
(high end)

Column height ∝ claimed speedup. Depending on workload, startup is reported as 2.5× to 50× faster than namespace-based Docker.

Under concurrency (company figures)

50 concurrent: 67ms avg · P95 90ms · P99 137ms
100 concurrent on a 96-vCPU host: P99 under 200ms

What's still unproven

Independent third-party benchmarks remain scarce
Long-term operational reviews and real-environment reproducibility unconfirmed

How it stays fast and isolated

KVM + RustVMM microVM

→

Dedicated Guest OS kernel per agent

→

CoW + snapshot clone + resource pool

→

Hardware isolation at microVM speed

E2B-compatible SDK (Python & Go): existing code migrates by changing only the endpoint. ARM64 support, auto pause/resume, and a Terraform deployer included. Cited in production by MiniMax's Agentic RL training at hundreds of thousands of instances per minute.

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.