# Operator Sandbox

Operator runs attack simulation campaigns against a configured target. It starts
Attack Harness, provides its permitted tools, applies limits, and retains the
inputs, observations, and results needed to review what happened.

The same campaign workflow can also support operational simulations that evaluate
the agent's functional effectiveness. Your objectives can ask whether it completes
legitimate tasks correctly, responds consistently to equivalent requests, or handles
the modeled conditions reliably. Attack simulations retain their security focus;
operational simulations use task outcomes as their evidence criteria. A campaign
can include both when you explicitly supply both sets of objectives.

## How the sandbox works

Operator Sandbox enforces isolation through constrained tools and a completely
network-isolated Attack Harness container. The administrator configures the tools
before a campaign launches. Attack Harness cannot modify those configurations or
expand its own permissions.

For example, Attack Harness can submit an attack payload and injection through an
exposed tool. Operator delivers that submission to the configured target. The
target's IP address or hostname is already set in Operator's administrator
configuration; Attack Harness cannot choose or override that delivery destination
through its submission. Choosing permitted test content is different from
controlling where Operator sends it.

Attack Harness has no network access of its own. Its only means of communicating
outside the sandbox is through the tools Operator exposes. Operator performs the
permitted model and target operations on its behalf and retains their results.

## How a campaign progresses

1. Configure a model provider, compatible local Attack Harness image, and target.
2. Export capabilities and author a bundle from your objectives.
3. Validate the inputs and start the campaign.
4. Attack Harness proposes experiments through Operator's tools and adapts to feedback.
5. Operator closes execution and retains evidence for inspection and reporting.

Interceptor supplies a controlled target environment and supported state operations.
Operator also supports configured HTTP/HTTPS targets; those do not gain simulated
services or snapshot capabilities merely by connecting them.

## Start using Operator

Follow [setup](guides/setup.md), then [prepare and run a campaign](guides/campaigns.md).
Use [scenario bundles](guides/scenario-bundles.md) for input authoring and
[results and reports](guides/results-and-reports.md) to interpret outcomes.

The [platform walkthrough](https://docs.intrusive.ai/pages/guides/attack-simulation-walkthrough.html)
connects Operator to the support agent in Interceptor. Check integration contracts
when selecting the components to run together.
