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

# Browser extension

> Discover by NeuralTrust — a managed browser extension that finds shadow AI, records what employees send to third-party AI services, and blocks or masks prompts before they leave the browser.

**Discover by NeuralTrust** is a managed browser extension for the surface nothing
else reaches: employees using third-party AI in a browser. No SDK to add, no
gateway in the path, and no cooperation needed from the AI vendor.

It answers two questions an org usually cannot: **which AI services are people
actually using**, and **what are they pasting into them**.

<Note>
  This is not a [TrustGuard collector](/trustguard/concepts/collectors). Collectors
  call [`/v1/evaluate`](/trustguard/api/evaluate) with a `tgk_…` key and enforce a
  runtime policy. The extension is a separate integration: it reports discovery and
  activity to the AISPM and Data Plane APIs, and checks prompts against the Actions
  API. It is configured entirely through managed browser policy — there is no
  collector to create for it.
</Note>

## What it covers

| Surface                    | Monitor | Block | Redact |
| -------------------------- | :-----: | :---: | :----: |
| Prompt, before it is sent  |    ✅    |   ✅   |    ✅   |
| Attachments, before upload |    ✅    |   ✅   |    ➖   |
| Model response             |    ✅    |   ❌   |    ❌   |

**Use it when** you need visibility and control over employee use of AI services
you do not operate — ChatGPT, Claude, Gemini and whatever appears next quarter.
**Not when** you want to protect your own AI product: that is a
[TrustGuard collector](/trustguard/integrations/coverage), and the two are
complementary rather than alternatives.

**Limits.**

* **Managed browsers only.** A personal device, a personal browser profile, or
  the native desktop app of the same AI service is outside the perimeter. This is
  the defining boundary — plan for it rather than around it.
* **Model responses are recorded, not enforced.** They reach activity so you can
  see what came back, but no policy decision is applied to them.
* **Chrome and Edge only.** The extension ships as a single Chrome Manifest V3
  package. Edge is Chromium and loads it. **There is no Firefox build**, and
  Firefox's MV3 implementation differs enough that the same package is not
  expected to load.
* Site-level blocking through `declarativeNetRequest` is not driven by policy
  today; enforcement happens per prompt, at submit time.

## How it works

The extension runs three layers in the browser:

| Layer          | What it does                                                                                                               |
| -------------- | -------------------------------------------------------------------------------------------------------------------------- |
| **Injected**   | Page-context interceptors over `fetch`, `XHR` and `WebSocket`, plus a prompt extractor that knows the shape of each AI app |
| **Content**    | Policy bridge, DOM monitor and the toast notifications the user sees                                                       |
| **Background** | Service worker: identity, config and policy sync, event buffering, and the moderation call                                 |

When a user submits a prompt, the extension extracts it, sends it for evaluation,
and acts on the verdict **before the request leaves the browser**:

* **Allowed** — the request proceeds and the interaction is recorded.
* **Masked** — sensitive spans are replaced and the masked text is submitted
  instead. When the PII mode is set to enforce, the same finding blocks instead
  of masking.
* **Blocked** — the request never leaves; the user gets an *Access Blocked* page
  with your `customBlockMessage`, your support email and a link to your IT
  portal, so the outcome is a route to permission rather than a dead end.

Attachments are checked on the same path before upload.

## Identity

The extension attributes activity to a real person rather than a device.

`userId` and `userEmail` can be pushed through managed policy. Where you want the
browser to establish identity itself, enable **Microsoft Entra ID** sign-in with
`entraLoginEnabled`, `entraTenantId` and `entraClientId`. Sessions are grouped
automatically, so no per-site configuration is needed.

## Deploy

Push the extension to managed endpoints with Chrome or Edge enterprise policy —
MDM, GPO or Intune. Force-install it; an extension that is merely allowed is not
a control.

Configuration arrives through the browser's **managed storage**, not through a
console. Four keys are required:

| Key                                    | Required | What it is                                            |
| -------------------------------------- | :------: | ----------------------------------------------------- |
| `teamId`                               |     ✅    | Your organization identifier                          |
| `integrationId`                        |     ✅    | The shadow-extension integration identifier           |
| `aispmEndpoint` · `aispmToken`         |     ✅    | AISPM API — discovery and inventory                   |
| `dataPlaneEndpoint` · `dataPlaneToken` |          | Data Plane API for activity ingestion (JWT bearer)    |
| `actionsEndpoint` · `actionsApiKey`    |          | Actions API for prompt checks, sent as `X-TG-API-Key` |

Without the Actions keys the extension observes and reports but does not enforce,
which is a reasonable first phase.

### Behavior

| Key                                                 | What it controls                                   |
| --------------------------------------------------- | -------------------------------------------------- |
| `syncInterval`                                      | How often buffered activity is flushed, in minutes |
| `policyRefreshInterval`                             | How often policy is re-fetched, in minutes         |
| `bufferSize` · `offlineBufferSize`                  | Events held before a flush, and while offline      |
| `allowUserOverride`                                 | Whether users may override certain policies        |
| `notificationsEnabled`                              | Desktop notifications                              |
| `customBlockMessage` · `supportEmail` · `portalUrl` | What the *Access Blocked* page tells the user      |
| `debugMode`                                         | Verbose logging for troubleshooting                |

### Rollout

1. Force-install the extension and push `teamId`, `integrationId` and the AISPM
   keys. Let discovery run and look at what appears — most orgs find services
   they did not know were in use.
2. Add the Data Plane keys so interactions land in activity.
3. Add the Actions keys with a report-oriented policy, review findings over a
   representative period, then enable blocking and PII enforcement.
4. Set `customBlockMessage`, `supportEmail` and `portalUrl` **before** enforcing.
   A block with no route to permission turns into a support ticket, or into
   someone switching to their phone.