# MECHANICS — How the sandboxes work.

> Operator Sandbox runs and governs the AI-powered tester. Interceptor Sandbox runs the target agent with declared dependencies and controlled injection points. Choose either sandbox on the website to see what runs inside it, how it communicates, and what gets recorded.

- Status: documented product mechanics; choose documentation matching your installed release.
- Last reviewed: 2026-11-18
- Intended audience: AI security practitioners, red teams, and application builders.
- Related products: Operator Sandbox, Attack Harness, Interceptor Sandbox.
- Canonical URLs: [website section](https://intrusive.ai/#mechanics), [this document](https://intrusive.ai/content/mechanics.md), [platform documentation](https://docs.intrusive.ai/pages/overview.html).

## Operator Sandbox: the testing side

Operator runs the AI-powered Attack Harness inside a network-isolated container. The trusted host controls every permitted connection to a model or target.

1. **Define the investigation.** Supply objectives and optional scenarios. Operator checks the inputs against target capabilities, configured permissions, and limits.
2. **Start the isolated tester.** Attack Harness plans experiments and adapts to feedback using a tool catalog fixed before the campaign starts.
3. **Broker the permitted actions.** Operator makes model and target requests on the harness's behalf. Credentials and target destinations stay under host control.
4. **Retain the results.** Review the accepted inputs, attempted actions, observations, and assessment together, including missing evidence and untested work.

**Boundary:** Attack Harness cannot change its tool configuration or choose a different target destination. A payload supplied through a permitted tool is different from authority to choose where Operator delivers it. The administrator's target profile grants permission; capability exports describe the available interface. Model-generated instructions cannot expand that permission.

Operator controls the campaign. Attack Harness performs adaptive testing within that control. Interceptor supplies the target environment when configured as Operator's target. Operator also supports configured HTTP/HTTPS targets; connecting those does not give them synthetic dependencies or snapshots.

## Interceptor Sandbox: the application side

Interceptor runs your agent in a container with the services, files, identities, and data declared for its environment. Observe how it behaves when those inputs change.

1. **Describe the agent's world.** An environment blueprint defines how to invoke the agent and configures its dependencies and starting data.
2. **Run against controlled dependencies.** Provide supported HTTP, MCP, PostgreSQL/pgvector, and file surfaces. Real model calls and declared HTTPS seams use configured host-side connections.
3. **Place payloads at the boundary.** Use the injection API to alter supported model, tool, service, or database responses, or declared mutable environment state.
4. **Observe and compare behavior.** Inspect recorded interactions and state changes. Where snapshot prerequisites are met, restore declared logical state to compare experiments.

**Boundary:** Declared connections do not grant general network access. Snapshots do not capture arbitrary process memory or remote service state. An undeclared dependency does not automatically become a working simulation.

Response injections change content in transit; environment-state injections change declared mutable session data. Placement, selection, lifetime, and permitted operations depend on the advertised capabilities and policy. Removing an injection definition does not undo a state mutation already applied. Payload placement, delivery evidence, an attempted tool action, and successful attack effects are different observations.

## Evidence and limits

Operator retains campaign inputs, operations, observations and conclusions. Interceptor records target interactions, available state changes and native evidence. Review those records together; the model's assessment alone is not proof of an effect. Missing observations remain uncertainty, and a completed campaign does not establish that all objectives were tested or succeeded.

Supported snapshots preserve declared logical state under the application's documented prerequisites. They do not promise deterministic model output. Stopping, reporting, exporting and purging are separate operations; administrators control retention in their installations. Synthetic testing still requires authorized targets and reviewed real external connections.

The miniature globes on the website are static conceptual diagrams. They show the isolated tester and contained target, their permitted connections and representative dependencies; they are not live telemetry or a complete capability inventory.

## First-party sources

These explanations are based on the product documentation in the INTRUSIVE.AI documentation collection, rather than inferred from the drawings:

- [Operator Sandbox overview](https://docs.intrusive.ai/pages/versions/operator_sandbox/0.1.5/overview.html): isolation, configured tools and host-owned destinations.
- [Prepare and run a campaign](https://docs.intrusive.ai/pages/versions/operator_sandbox/0.1.5/guides/campaigns.html): admission, capabilities, execution and limits.
- [Results and reports](https://docs.intrusive.ai/pages/versions/operator_sandbox/0.1.5/guides/results-and-reports.html): observations, conclusions, coverage and retention.
- [Attack Harness overview](https://docs.intrusive.ai/pages/versions/attack_harness/0.1.5/overview.html): adaptive experiments within Operator's permissions.
- [Interceptor Sandbox overview](https://docs.intrusive.ai/pages/versions/interceptor_sandbox/0.1.5/overview.html): synthetic dependencies, injection surfaces and snapshot boundaries.
- [Environment blueprints](https://docs.intrusive.ai/pages/versions/interceptor_sandbox/0.1.5/guides/environment-blueprints.html), [services](https://docs.intrusive.ai/pages/versions/interceptor_sandbox/0.1.5/guides/services.html), and [sessions and evidence](https://docs.intrusive.ai/pages/versions/interceptor_sandbox/0.1.5/guides/sessions-and-evidence.html): environment configuration and observation paths.

These are documentation references, not a claim of a newly executed product qualification run. Select documentation matching installed component versions when integrating or operating the products.
