> ## Documentation Index
> Fetch the complete documentation index at: https://docs.beyondguard.ai/llms.txt
> Use this file to discover all available pages before exploring further.

# Sandbox: Security-Scan Models Before Production

> Run externally sourced AI models through malware, format-integrity, and static code analysis in an isolated environment — with live progress and a downloadable report — before they reach production.

The **Sandbox** is BeyondGuard's pre-production control gate. Rather than deploying a model from Hugging Face, an open-source repository, or your own Git directly into production, you first run it through a security scan in an **isolated environment** — before any policy opens it to traffic.

## Starting a run

From **Start New Sandbox Run**, each scan is given a run name in the **Basic Information** section, and you configure:

* **Source** — `Huggingface`, `Custom`, or `Git`, so models from different distribution platforms go through the same standard process.
* **Model Configuration** — the model **type**, the **Model URL**, and an optional **HF Token** for pulling from gated or private repositories.
* **Security Checks** — select which analyses to run, customizing the scope of the scan to your needs.

The scan covers **malware checks**, **file/format integrity checks**, and **static analysis** of the scripts and code contained within the model.

## Live monitoring and results

Once started, the entire analysis is observable in real time on the **Model Information** screen:

* **Live Data Stream** — shows the stages (Download → Checks → Report) as percentages with progress bars; the download, antivirus scan, format-integrity check, and script/code analysis steps can each be watched separately.
* **Event Stream** — lists the result of each check chronologically with timestamps (files scanned/skipped, findings detected), marks each check **Success** or **Failed**, and expands to show duration and sub-details.

When the run completes, a detailed report of all findings is exported with **Report Download**.

<Tip>
  The Sandbox lets you verify end to end, from one screen, whether an externally sourced model contains malware, corrupted formats, or risky code before it is brought into production — reducing manual review and shortening time to production.
</Tip>

## Related

<CardGroup cols={2}>
  <Card title="Model Library" icon="boxes-stacked" href="/platform/model-library">
    Where scanned models are registered.
  </Card>

  <Card title="Red Teaming" icon="crosshairs" href="/guides/red-teaming">
    Behavioral attack testing, complementary to static scanning.
  </Card>

  <Card title="Shadow AI" icon="eye" href="/platform/shadow-ai">
    Catch models and services in use without approval.
  </Card>

  <Card title="Platform Overview" icon="grid-2" href="/platform/overview">
    Where the Sandbox fits in the control plane.
  </Card>
</CardGroup>


This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.