Browser Extension

The Netzilo browser extension governs AI use inside the browser. It runs the same governance engine as the Netzilo client and the Netzilo AI Gateway: the same guardrails, the same behavior graph and the same AI events. A guardrail written for devices applies in the browser unchanged.

Use it when people use web AI apps such as ChatGPT or claude.ai, on managed devices and on personal ones. On Linux it is the only way to govern AI in the browser.

You give the extension its policy under Edge > Governance > Filters, with an endpoint filter whose Operating Systems include Browser. Its verdicts appear under Activity > Events.

Watch · 1:47

Govern ChatGPT and Claude in the browser

Redact a customer ticket and block a private key in ChatGPT and claude.ai, with only the extension installed.

All how-to videos →YouTube ↗

How it works

  1. The user signs in to the extension once, at the Workplace.
  2. The extension reads the endpoint filters whose Operating Systems include Browser and whose groups contain the user, with their guardrails.
  3. When a page sends a prompt or receives an answer, the extension evaluates it against those guardrails before it leaves the browser or is shown on the page.
  4. Each verdict becomes an AI event in Activity > Events, and the tab's activity is uploaded to the user's behavior graph.

Guardrails are data from the Netzilo Server. A new or changed guardrail reaches the extension without a new extension release.

What it governs

The extension sees the requests a page makes: fetch, XHR, WebSocket and EventSource.

TrafficExamplesGoverned as
Web AI appsChatGPT (signed in and logged out), claude.ai, Gemini, Perplexityprompts and answers, including streamed answers
API calls from a pagea web app or an in-page script calling the OpenAI, Anthropic or Google AI APIs, or other LLM hosts the engine knows, including a local Ollama on localhost:11434prompts and answers
MCP over WebMCPa page that offers tools through the browser's WebMCP interfacetool calls and results
The tab's HTTP requestsevery request the tab makesa network log of host, path, method and content type, never bodies
Conversation historyChatGPT, claude.ai and Gemini reloading a conversationgoverned again when shown, so a redaction also applies after a reload

It governs nothing outside the browser and never inspects the Netzilo client's traffic.

A page that offers its own tools to an AI agent through WebMCP gets the same treatment as an MCP server: each tool call and each result passes the guardrails, so a result can come back redacted and a call can be refused before the tool runs.

Watch · 1:45

Govern the tools a page gives to AI agents

A web app offers WebMCP tools to its assistant: results are redacted and a refund is stopped before it runs.

All how-to videos →YouTube ↗

Browsers

BrowserPackageNotes
ChromeChrome Web Store, ID kdcpnkonhjcknlkjmjapclhpoamlckjinetwork log available
EdgeEdge Add-ons, ID efflopncanncclcfnddajplbnmhoebmlnetwork log available
Firefox 128 or laternetzilo-secure-browsernetwork log available
Safarithe Netzilo Secure Browser app from the App Store (ID 6740890085); the Safari extension inside it is named Netzilo Browserno network log
Netzilo Enterprise Browserbundledsee Enterprise Browser

Safari offers no network log, so guardrails that declare requires: http are skipped there. Rules on prompts, answers and tools work in Safari as everywhere else. See Detection Rules.

Before you start

  • Guardrails under Edge > Governance > Guardrails, see Guardrails.
  • The users in a group, for example Employees. The browser is not a peer: the extension matches filters by the user's groups.

Step 1: Bind guardrails to the browser

  1. Go to Edge > Governance > Filters and click Add Filter, or open an existing endpoint filter.
  2. Set Filter Type to Endpoint.
  3. Under Operating Systems, select Browser. You can also keep desktop systems in the same filter, for example Windows and Browser; the filter then serves both.
  4. Under groups, add the users' groups. The group All matches everyone.
  5. Add the guardrails, name the filter and save.

The Create New Filter dialog on the Targets tab with Filter Type Endpoint, the group Developers, and the operating systems Windows, Mac and Browser selected

The Agents list is not used by the extension: process paths mean nothing in a browser. Leave it empty on a browser-only filter. On a mixed filter it still steers desktop processes on Windows and macOS.

A desktop operating system never matches the extension, and a filter with only Browser never reaches a desktop client. When several filters match a user, their guardrails are merged into one rule set.

Step 2: Install the extension

  • Force-install with browser policy. Chrome: add kdcpnkonhjcknlkjmjapclhpoamlckji;https://clients2.google.com/service/update2/crx to ExtensionInstallForcelist. Edge: add efflopncanncclcfnddajplbnmhoebml;https://edge.microsoft.com/extensionwebstorebase/v1/crx.
  • Let the Netzilo client install it. When an Enterprise Browser Extension profile applies to a device, the client writes the force-install policy for Chrome and Edge itself, and removes it when the client service is uninstalled. See Browser Extension profile.
  • Safari: users install the Netzilo Secure Browser app, then enable Netzilo Browser under Safari Settings > Extensions and allow it on all websites.

Step 3: Users sign in

The user opens the extension's popup and clicks Login. This opens the Workplace, where the user picks the tenant and signs in. The popup then shows Protection: Active.

The icon buttons at the top of the popup are Workplace, pause, help and refresh. The Workplace button (the grid icon) opens the extension's Workplace page in a new tab: the same Workplace as in the dashboard. Signed out, that page shows the workplace picker: Default Workplace, the workplaces the user added, and Add Workplace, which asks for the Workplace URL. Click Connect next to one to sign in. In the Netzilo Enterprise Browser the same page is the new tab.

After you change a filter or a guardrail, the user can press Refresh in the popup to load it at once. The popup's Debug tab shows the rules version and the number of filters, so you can confirm the change arrived.

When the Netzilo client also runs on the device

Only one side governs a browser, so nothing is evaluated twice. The Netzilo client takes over the browser on Windows and macOS when both of these are true:

  • the browser's executable is in the Agents list of an endpoint filter for the device, and
  • the AI sites (chatgpt.com, claude.ai, gemini.google.com) are inspected by the client.

The extension then stands down after the Netzilo Server confirms that the user has a connected device. The popup says "the desktop client governs this browser", and events come from the client instead of the extension. On Linux the client does not steer browser traffic, so the extension always governs the browser.

Fail-open and fail-closed

While the engine cannot judge a request (it is starting, the Netzilo Server is unreachable, or the request cannot be read), the extension lets traffic through by default. To stop such requests instead, set the managed browser policy fail_mode to closed for the extension, with the same browser management you use for force-install. Fail-closed applies only to AI endpoints; every other site always loads. There is no dashboard setting for it.

Route a page's AI calls through the AI Gateway

A web app that calls a model API from the browser, with a key in its code and a model of its own choosing, can be sent through your AI Gateway without changing the app. Bind a guardrail with the replacemodel action and a base_url to the browser filter:

title: Route web apps through the AI gateway
logsource:
  product: ai_agent
  category: llm_request
detection:
  anthropic:
    provider|re: '^anthropic$'
  condition: anthropic
action: replacemodel
model: claude-haiku-4-5
base_url: https://aigw.example.com/anthropic
  • The extension relays the call, as the user. It sends the request to the gateway itself, with the signed-in user's Netzilo credential. The page's own API key, cookies and authorization header are not sent. The gateway then applies that user's filters, models and limits.
  • model is the model the gateway serves. Leave it out to keep the page's choice.
  • One session, seen twice. The call is one AI session under Edge > Sessions, seen in the browser and at the gateway, and each relayed call is an AI redirected event with the model the page asked for and the model that answered.
  • What is routed. API calls from pages are routed; web AI apps such as ChatGPT keep their own backends. base_url must use https, or http on localhost. A gateway reached as a published application works, so the route can name an address on your own domain.

Watch · 1:59

Bring web apps' AI calls under the gateway

One routing guardrail sends a web app’s model calls through the AI Gateway as the signed-in user, on an approved model.

All how-to videos →YouTube ↗

What users see

VerdictIn the page
RedactThe matched text is replaced by the rule's replacement, for example [CARD-REDACTED], in the prompt as it is sent and in the answer as it is shown
Block on a promptChatGPT shows an error with the rule's reason. claude.ai and Gemini answer with Netzilo Guardrail: Blocked by rule: <rule name>
DetectNothing in the page; an AI detected event is recorded
Streamed answersText is released about 1 KB behind the model, so that a redaction applies before it is shown. Long answers appear with a short delay; short answers appear at once at the end
Reload of a signed-in ChatGPT conversationThe messages are hidden briefly until the governed history has loaded, then shown with redactions applied
Isolation (response action)The extension's notification on the page, Response action: AI access is isolated by an administrator. with the reason and a reference, see Response actions

The extension never answers with a block page; refusals appear inside the chat.

Response actions

From version 5.2.578, the extension applies response actions. When the user, one of their groups, an AI app they use, or all AI activity is isolated, whether by an administrator or by Netzilo on its own, the extension refuses the user's AI activity in the browser: prompts and answers in web AI apps, AI API calls from pages and WebMCP tool calls. A stream in progress is cut.

  • The extension checks for response actions every 30 seconds and at sign-in, so an isolation or a lift takes effect in the browser within about 30 seconds.
  • It keeps the actions it holds, so an isolation also holds while the Netzilo Server cannot be reached.
  • A web app shows only its own generic error for a refused call, so the extension shows the refusal as a notification on the page, with the reason and a reference.
  • Guardrails that read or apply response actions (requires: [..., controls]), such as Cruise Control Inspection and Containment, load in the extension from 5.2.578 on.

Older extensions do not apply response actions and skip those guardrails. Update the extension, for example by force-installing the current version through the Browser Extension profile.

Where the results show

Every verdict is an AI event under Activity > Events, in the category AI Edge. The row reads <site> in the browser for <user>, followed by what happened, the rule and the filter.

The Activity Events page filtered to AI redacted, showing rows that read chatgpt.com in the browser for alice@example.com, prompt to gpt-6 redacted by rule PII in prompt (redact) by filter Everyone — browser, with the redacted prompt text below

  • Each browser tab on a site is a session of its own, under a browser session. The first time a tab session shows behavior, Activity records a Session snapshot entry, "Session snapshot is available for <site> in the browser", which opens the session's behavior graph. See Behavior Graph & Kill Chains.
  • Regular users see only their own browser sessions; administrators see all of them.
  • Conversations with AI apps such as ChatGPT and claude.ai are recorded as AI sessions of the signed-in user, with their turns, tool calls and token counts (no cost: pricing is only known at the AI Gateway). They are listed under Edge > Sessions next to the sessions of devices and the AI Gateway, and show Under Isolation or Under Inspection once the browser applies a response action.
  • The AI Agent Activity report counts browser events.
  • Uploaded observations are kept for 7 days. Events are kept like every other activity event.

Troubleshooting

Ask the user for the popup's Debug tab first. It shows the state, Protection, whether the extension stood down, the rules version, the number of filters, the rules loaded, failed and skipped by name, and the browser's capabilities.

Prompts pass and no events appear

The user is not signed in, or no enabled endpoint filter with Browser includes one of the user's groups (the popup shows 0 filters). Sign in from the popup, add Browser and the user's group to the filter, then press Refresh.

The Workplace page shows This address is not a Netzilo workplace

The address entered under Add Workplace is not the Workplace address of a Netzilo Server, or the browser cannot reach it. Enter the address the user signs in to the Workplace at, starting with https://.

The popup shows Protection: Inactive

There is no sign-in, the user's filters are empty, or the extension stood down. The popup names the reason.

The extension stood down, but nobody governs the AI sites

The browser's executable is in the filter's Agents list, but the client does not inspect the AI sites. Remove the browser from Agents so that the extension governs it.

A rule is listed as failed in the popup

The guardrail does not compile, for example an invalid regular expression. Fix it under Edge > Governance > Guardrails, save, and press Refresh.

A rule is listed as skipped in Safari

The rule needs the network log (requires: http), which Safari does not offer. It works in Chrome, Edge and Firefox.

Events show the source mitm instead of extension

The Netzilo client governs this browser. This is expected on Windows and macOS when the browser is in the filter's Agents list.

Works in Chrome, nothing in Safari

The Netzilo Browser extension is not enabled, or not allowed on all websites, under Safari Settings > Extensions.

An isolated user can still use AI in the browser

The extension is older than 5.2.578, or has not checked for response actions yet (it checks every 30 seconds and at sign-in). Check the extension's version on the browser's extensions page and update it.

Unmapped counters rise in the Debug tab

The engine could not read traffic on an AI site, usually because the site changed. Report the site and the counters to Netzilo support. Rules on HTTP requests still apply.

Limits to know

  • A logged-out ChatGPT conversation reopened by a reload is not redacted, because the site renders it on its server. New prompts and answers in it are governed.
  • Safari has no network log: rules on HTTP requests are skipped there.
  • Fail-closed covers AI endpoints only, and it is a browser policy, not a dashboard setting. The default is fail-open.
  • The browser is not a peer: it has no posture checks and no network policies, and filters reach it through the user's groups only.
  • One side governs a browser at a time. When the client inspects the browser but not the AI sites, nobody governs those sites until the filter's Agents list is corrected.