Verdicts · new leads

New sandbox leads

A compact reading list for projects that surfaced after the main research pass. These are direct links and starting points, not fresh ratings or an endorsement.

How to use this page

Start with a project’s threat model, host requirements, network defaults and persistence model. A fast demo is useful evidence, but it is not evidence of containment. Links go to the project itself or its documentation; there are no affiliate or synthetic referral parameters.

E2B is the lead worth a dedicated look

E2B core ↗

E2B presents a developer platform for running code and tools in cloud sandboxes. It is the useful starting point if the goal is an API-first sandbox service and a separate deployment evaluation.

Read the docs ↗ · Project site ↗

Other projects to put on the evaluation board

ZeroBoot ↗

Copy-on-write VM forking aimed at fast, concurrent agent sandboxes.

VMTREE ↗

Self-hosted ephemeral VM provisioning; relevant for a host-side pool or template workflow.

Microsoft Quicksand ↗

QEMU-based VM automation with a Python API, desktop-oriented controls and checkpoint-oriented workflows.

CubeSandbox ↗

A multi-tenant, lightweight sandbox project for agent workloads.

Matchlock ↗

Linux sandboxing focused on securing agent workloads.

stereOS ↗

A Linux system positioned as hardened and purpose-built for AI agents.

Where this fits the local experiment

For a Windows-first desktop product, separate the user-facing orchestration experience from the workload boundary. A practical early comparison is: an Incus container for dense, trusted task isolation; an Incus VM or a KVM microVM for a separate kernel; and an API-hosted model such as E2B where the product needs remote execution. The Sandbox Stacks chapter explains the boundary layers that still need to sit around each option.