Tencent open-sources Cube Sandbox
- Tencent published Cube Sandbox, a RustVMM-based runtime for running AI agents with hardware isolation. - The runtime advertises sub-60ms start times, low memory usage, and hardware-backed isolation for agent workloads. - The project aims to make agent-scale deployments faster and safer for teams building production AI services (x.com).
Tencent Cloud has open-sourced CubeSandbox, a RustVMM-based microVM runtime for running AI agents with sub-60ms cold starts and hardware-backed isolation. (github.com) Tencent published the CubeSandbox repository and docs this week and released the code under the Apache License 2.0; the repo and license are on GitHub. (aihola.com) The project builds on RustVMM and KVM, advertises cold-starts under 60 milliseconds, claims per-instance memory overhead under 5 MB, and supports single-node or multi-node cluster deployments. (github.com) AI agents commonly run short-lived, untrusted code inside sandboxes; E2B and microVMs like Firecracker set the current technical baseline for isolated, fast-starting sandboxes. (e2b.dev) CubeSandbox natively implements the E2B SDK interface and is pitched as a drop-in, self-hosted alternative to E2B cloud sandboxes, letting teams swap endpoints without changing application code. (docs.cubesandbox.ai) The codebase wires together six components — CubeAPI (gateway), CubeMaster (orchestrator), Cubelet (node manager), CubeVS (eBPF virtual switch), CubeRuntime, and a hypervisor shim — to deliver snapshots, CoW memory reuse, and network isolation. (chooseai.net) Tencent’s README and blog numbers are measured in their lab: average sandbox creation under 60ms at small concurrency, with reported P95 around 90ms and P99 near 137ms in some tests; those figures are vendor-provided, not independent benchmarks. (aihola.com) The full source, quickstart examples (including OpenAI Agents SDK integration), and documentation are available on the GitHub repo and the CubeSandbox docs site for engineers to test and contribute. (github.com)
