In this piece
Alibaba OpenSandbox Review: What to Check Before Self-Hosting
You want an AI agent to run code on your infrastructure. Before giving it a customer credential, check the runtime boundary and what survives sandbox reuse.
OpenSandbox is an open-source platform for running AI-agent workloads on infrastructure you control. Its lifecycle API connects applications to Docker or Kubernetes deployments, including optional microVM infrastructure. We would shortlist it if infrastructure control justifies operating and testing the platform.
Research status: 11 November 2026. We checked the 1.2.0 source and official documentation alongside upstream reports. We did not deploy or benchmark OpenSandbox, or reproduce the reported security issues. The proposed tests below are acceptance criteria.
What is Alibaba OpenSandbox, and who maintains it now?
OpenSandbox supplies sandbox lifecycle and execution services: create an environment, run commands, manage files, obtain endpoints and terminate the workload. Alibaba contributed the project to the Agentic AI Foundation, announced on 22 October 2026. The current repository is opensandbox-group/OpenSandbox, licensed under Apache 2.0.
The project offers Python, JavaScript/TypeScript, Kotlin/Java, Go and C#/.NET SDKs, plus CLI and MCP interfaces. Official examples cover coding agents, browser automation and desktops. Test the complete combination of agent and deployment versions before using customer data.
The architecture separates the lifecycle server from execution and networking. Your application still decides what the agent may attempt, when a person must approve it and how success is verified. Our agent harness engineering guide explains that application layer.
Which OpenSandbox features are stable in November 2026?
At our check, 1.2.0 was the latest stable release, published on 27 October. Its release notes add opt-in FastSandbox auto-pause and wake-on-access, Credential Vault integration for fast sandboxes, and OCI-backed snapshot workflows. Each depends on its backend and configuration. For reproducibility, the runtime details below use the source at release-1.2.0, commit 62a64f0.
| Surface | Reviewed baseline | Deployment consequence |
|---|---|---|
| Server, CLI and published SDK packages | 1.2.0 | Pin a coherent release instead of combining arbitrary package versions. |
| Platform component images | release-1.2.0 | Resolve the intended image digest and retain the release BOM. |
| FastSandbox idle automation | Included in 1.2.0, disabled by default | Enable the backend, infrastructure controls and per-sandbox policy together. |
| Credential Vault Exchange | In 1.2.1-rc.2, a prerelease | Its credential-exchange API requires a newer release than 1.2.0. |
| Workload images, such as Code Interpreter | Versioned separately | Pin and review the workload image separately. |
The 11 November release candidate publishes images while SDK and CLI packages wait for the stable bump. The SDK documentation warns that it follows the default branch.
For the next upgrade, include a Kubernetes controller restart with active sandboxes and a lower pool size with idle capacity. Verify that active work survives and unused capacity is released. Run these checks on the pinned images and packages you intend to deploy.
The versioning policy now aligns platform releases under one version, while workload template images remain separate. It also describes latest-line support without LTS. Budget for upgrades.
Should you use Docker, Kubernetes or FastSandbox?
Use Docker to validate a local integration; choose Kubernetes when you need its deployment and tenant controls. Consider FastSandbox when template-backed microVMs and checkpoint workflows serve the workload. Verify networking and state recovery on your selected runtime.
| Deployment | Useful starting point | What to verify |
|---|---|---|
Docker with ordinary runc | Trusted development and SDK integration | Host exposure, mounts, API authentication and cleanup; ordinary containers share the host kernel. |
| Docker or Kubernetes with gVisor | Workloads compatible with a user-space kernel | Syscall compatibility and an alternative to OpenSandbox’s incompatible NAT-based egress sidecar. |
| Kubernetes with Kata | Container workflows needing a guest-kernel boundary | Installed RuntimeClasses, capacity, image compatibility and the selected networking controls. |
| FastSandbox with Firecracker | Template-backed microVM and checkpoint workflows | KVM-capable infrastructure, templates, storage, shared platform components and per-VM policy enforcement. |
The secure-runtime guide distinguishes gVisor’s user-space kernel from Kata VM runtimes. Firecracker is also available through the separate FastSandbox architecture; it is not automatically enabled by a Docker quick start. For host and operating-system decisions, use our Linux infrastructure guide for AI agents.
Does every OpenSandbox get network isolation?
OpenSandbox’s network controls depend on the chosen profile and configuration. For the classic sidecar path, verify that the egress component is attached and enforces your policy. Inbound endpoint authentication is a separate control. Test denied access separately.
The server configuration guide requires an API key for authenticated single-tenant operation. Secure Access protects inbound sandbox endpoints through a Kubernetes ingress gateway; Docker rejects secureAccess: true. Neither setting replaces outbound policy.
For a supplied policy, the 1.2.0 sidecar parser defaults a missing defaultAction to deny. FastSandbox with no networkPolicy explicitly installs a default-allow policy. Specify the policy explicitly; omission of the policy and omission of one field differ.
The network-isolation guide identifies these constraints:
- Platform policy: operator-controlled
deny.alwaysrules take precedence over user rules in the egress component. A policy supplied by the workload creator is a different trust boundary. - Pre-warmed classic pools:
extensions.poolRefcannot be combined with per-requestnetworkPolicy. Configure controls in the Pool pod template before allocation, or choose a non-pooled path. This restriction belongs to the classic Kubernetes service; FastSandbox maps per-request policy andpoolReftogether. - Runtime compatibility: the sidecar’s NAT interception is incompatible with gVisor’s netstack. Transparent Istio/Envoy interception in the same pod network namespace is also unsupported.
Test policy changes against existing connections as well as new ones. Upstream report #2329 describes established TCP sessions surviving a tightened policy on server 0.2.1 and egress 1.1.4. We have not reproduced it on 1.2.0. Verify the stop procedure. If a policy update leaves a connection usable, your incident response still needs to cut traffic and revoke the credential.
Does Credential Vault keep real API keys out of agent workloads?
Credential Vault is designed to broker outbound credentials so the workload can use placeholders instead of holding the real key. Verify destination binding before supplying customer secrets. Keeping a key out of guest files does not establish where the proxy sends it.
For classic sidecar deployments, the Vault guide requires Credential Proxy with a network policy and dns+nft enforcement. It recommends defaultAction="deny"; default-allow remains temporarily accepted with a warning. Bind credentials narrowly to the intended destination, method and path. A broad binding can still authorize an expensive or destructive API operation. The 1.2.0 implementation selects a unique highest-precedence binding and rejects a tie; binding order alone is not your access policy.
An open upstream report describes credential selection using HTTP Host without a matching TLS-peer check. The proposed fix #1759 was unmerged at our check. We have not reproduced this report; profiles differ. Require verified binding behavior on the deployed image before using customer credentials.
Are OpenSandbox pools isolated by tenant?
Classic Pool management APIs in OpenSandbox 1.2.0 are shared, even when tenant keys scope sandbox lifecycle operations. The multi-tenancy guide states that /pools operates in the server’s configured default namespace. Treat those pools as shared resources unless you establish a separate deployment boundary.
Docker does not support this namespace-based tenant model. On Kubernetes, resource quotas, default limits and network enforcement still need configuration. Test two tenant identities against list, read, delete and pool routes, then attempt traffic between workloads.
What happens to state and cost when a sandbox pauses?
A FastSandbox pause checkpoints microVM state and releases its execution capacity. Storage and shared-platform costs remain; other backends have different pause behavior. Test the resume path. A restored file is insufficient if the agent loses its task state or credentials behave differently after recovery.
The FastSandbox checkpoint documentation describes filesystem, memory and VM-state artifacts. Its idle-automation design is opt-in and keeps the expiry clock running while paused. Ingress inactivity can also miss background work, which is why an application activity probe may be needed.
Client cleanup and remote termination are separate. The TypeScript SDK example calls kill() to terminate the sandbox and close() to release local resources. Confirm remote termination and handle cleanup failures.
Can you verify the 1.2.0 artifacts before deployment?
You can pin image identities, but the current official documentation identifies missing signing and provenance coverage for examined 1.2.0 artifacts. The current default-branch release-verification guide distinguishes the legacy signed workflows from umbrella releases and points to issue #2463.
Record the release tag, package versions, image digests, workload-image digest and actual verification outputs in a deployment manifest. The release BOM’s build commit can differ from the tag commit. A matching digest proves which bytes you selected; it does not identify who built them. If verified provenance is a mandatory launch condition, keep that condition open until the artifact satisfies it.
When is OpenSandbox worth self-hosting?
We would choose OpenSandbox when control over infrastructure, runtime choice or integration with private systems justifies a team to operate it. A managed sandbox is worth comparing when your team needs an external operator for the execution fleet. Check the provider’s isolation and retention commitments, including available regions and support.
| Decision | Reason to shortlist OpenSandbox | Reason to compare a managed service |
|---|---|---|
| Deployment control | You need a specific cluster or runtime. | Available regions and runtime controls satisfy the workload. |
| Operating capacity | You have an owner for patching and incidents. | Fleet operation would displace product work. |
| Workload fit | Custom images and platform integration matter. | A standard execution API covers the requirement. |
| Economics | Measured usage can justify shared capacity and operator time. | Uncertain volume favors testing a usage-based service first. |
Compute, idle capacity, storage, networking, logs, upgrades and on-call work remain costs. Measure cost per accepted agent action with the same task set and quality threshold on both paths. Avoid comparing a vendor’s warm-start benchmark with your own cold deployment.
What should a 30-day OpenSandbox pilot prove?
A useful OpenSandbox pilot should produce a deployment-specific acceptance record and an operating-cost comparison. The proposed schedule is not a delivery guarantee. Start with synthetic secrets. Choose a reversible workload, define pass criteria before testing and retain failed results alongside successful runs.
| Window | Decision to test | Evidence required |
|---|---|---|
| Days 1 to 5 | One release and runtime profile | Package and image manifest, workload definition, owners and explicit provenance status. |
| Days 6 to 10 | Endpoint and tenant access | Expected denials for unauthorized callers and cross-tenant requests, including Pool routes. |
| Days 11 to 15 | Egress and credential bindings | Allowed and denied requests, independent TLS-peer/Host checks, existing-connection behavior and a tested stop procedure. |
| Days 16 to 20 | Pause, resume and termination | Recovered task state, correct secret lifecycle, expiry behavior and confirmed remote cleanup. |
| Days 21 to 25 | Upgrade and recovery | A rollback rehearsal and traces that connect identity, workload, policy and operator actions. |
| Days 26 to 30 | Self-hosted versus managed baseline | Accepted-task rate, latency distribution, infrastructure cost, operator hours and unresolved launch conditions. |
Use our AI-agent sandbox containment checklist for the broader threat model. Keep mandatory conditions open while evidence is missing.
Our recommendation
Before customer workloads, require evidence for credential binding and tenant boundaries, plus a tested way to stop network access. Resolve any mandatory artifact-verification requirements. Choose a managed baseline if the platform effort cannot be justified by a concrete infrastructure requirement or measured operating costs.
Wavect’s AI product engineering service can help define that workload and its acceptance criteria. The Twinsoft AI case study provides adjacent AI delivery context, not evidence of an OpenSandbox deployment. Our external AI support versus an in-house team comparison helps assess operating capacity. Discuss an OpenSandbox deployment review with the runtime, tenant model and credential requirements you need to evaluate.
