GUIDE, read 12 Nov 2026
Browser agents act with the user's logged-in sessions, so control extension installs by policy and limit where agents may click
Two different things get called browser AI, and they need different controls
The first is an AI extension: add-on code that runs inside Chrome or Edge and reads or changes pages. Writing assistants, meeting summarizers, page-chat tools and coding helpers are typical. An extension is installed from a store, runs under the permissions in its manifest and can be managed by enterprise policy.
The second is an agentic browser, where the AI is built into the browser itself and can act on pages for the user: click, fill forms, move between tabs and carry out multi-step tasks. Perplexity's Comet is one example, and its site lists Mac, Windows, iOS and Android as the platforms (Perplexity, Comet page, read 12 Nov 2026, vendor claim). OpenAI's ChatGPT Atlas and Gemini in Chrome are two others. For those two, the vendor pages we tried returned an error to our fetch on 12 Nov 2026, so this guide makes no product-specific claim about their admin controls. Read each vendor's current enterprise documentation before you decide.
The split matters because the control points differ. For an extension, the control point is the browser's extension policy. For an agentic browser, the first control point is whether the browser may be installed at all, since it is a separate application. After that you control the sites it can reach and the accounts it can use.
Extension permissions define what the add-on can see
Chrome shows users a permission warning for powerful APIs. Google's developer documentation explains that these warnings describe the capabilities an API grants, and that some warnings are suppressed when a broader one is present: the warning for the tabs permission does not show when the extension also requests all URLs (Chrome for Developers, permission warnings page, read 12 Nov 2026). The practical reading for a security team is that an extension with access to all URLs can read and modify every page the user opens, including internal tools, webmail and admin consoles.
An AI extension usually needs that breadth, because it reads whatever page the user asks about. That is the core tension. The feature that makes the extension useful is the same access that makes a malicious or compromised update dangerous. A trusted extension that is sold to a new owner and updated can start sending page content to a new endpoint without the user doing anything.
- Review the manifest permissions for each AI extension: host permissions, tabs, cookies, clipboard, scripting, webRequest.
- Check where the extension sends data: the vendor's domain, a model provider, or a third party. Read the privacy policy, and compare it with traffic you can see.
- Check who publishes it and how often it updates. A change of publisher is an event worth an alert.
- Compare store listing claims with observed behavior on a test profile.
Chrome and Edge policy can allow, block and force extensions
Both browsers manage extensions through policy. Microsoft documents the ExtensionSettings policy for Edge, which can control settings such as the update URL and blocked permissions, and which can override other extension policies set elsewhere in Group Policy, including ExtensionAllowedTypes, ExtensionInstallBlocklist, ExtensionInstallForcelist, ExtensionInstallSources and ExtensionInstallAllowlist (Microsoft Learn, Edge extension management reference, read 12 Nov 2026). Google publishes the same policy for Chrome in its Chrome Enterprise help (Google, Configure ExtensionSettings policy, read 12 Nov 2026).
The strongest posture is a default deny: block all extensions, then allow the ones you have reviewed by extension ID. That posture stops an employee from installing a new AI helper from the store on a managed browser profile. A weaker posture, a blocklist of named AI extensions, fails quietly as new ones appear weekly.
Two details deserve attention. First, the ExtensionSettings policy can restrict the permissions an allowed extension may use, so you can allow a tool and still block, for example, a host permission you do not want it to hold. Second, Google's help page on Incognito notes that if you install extensions for data protection, you need to make sure users cannot open them in Incognito mode and escape those controls, and it points to the IncognitoModeAvailability policy (Google, Chrome Enterprise Help, read 12 Nov 2026). A control that disappears in a private window is not a control.
Agentic browsers add actions, and actions bring prompt injection into the browser
A chat sidebar reads a page and answers. An agent reads a page and then does something. That difference moves the risk. Text on a page, in an email or in a document can contain instructions aimed at the AI, and if the agent follows them, it acts with the user's logged-in sessions.
Anthropic's research post on prompt injection in browser use, published 24 Nov 2025, says it directly: browser use creates unique prompt injection risks, and no browser agent is immune. In that post, Anthropic reported a 1% attack success rate for its newest model against an internal adaptive attacker given 100 attempts per environment, and called 1% a significant improvement that still represents meaningful risk (Anthropic, prompt injection defenses post, read 12 Nov 2026). That is a vendor's measure of its own system, and it comes with the vendor's warning that the problem is not solved.
The OWASP guidance on excessive agency frames the same issue for any LLM application. It names three root causes: excessive functionality, excessive permissions and excessive autonomy (OWASP GenAI Security Project, LLM06 excessive agency, read 12 Nov 2026). A browser agent that can act on any site, with every session the user holds, with no confirmation step, has all three.
The sessions an agent inherits are the real exposure
A browser agent does not need your password. It runs in a profile where the user is already signed in to email, the SSO portal, the CRM, the cloud console and the code host. Whatever the user can do in those sessions, the agent can attempt.
This is why the browser profile is the unit to think about. A managed work profile with SSO sessions and a personal profile with a consumer AI account are different risks. If an agentic browser is installed next to the managed browser and the user signs in to work apps in it, the work sessions now live in an application that your extension policy, your DLP extension and your network controls may not cover.
Check three things on any agentic browser you evaluate: whether it can run in a mode with no access to signed-in sessions, whether it asks for confirmation before sensitive actions such as sending mail, buying or changing settings, and whether administrators can restrict the sites it may act on. Take the answers from current vendor documentation and then confirm them on a test machine.
Data leaves through paste, upload and page reads, and DLP can see some of it
Browser-based AI moves data in three ways: the user pastes text into a prompt, the user uploads a file, or the AI reads the page. Microsoft's Purview documentation states that DLP policies can warn or block users from sharing sensitive information with third-party generative AI sites accessed through a browser, for example stopping a user from pasting credit card numbers into ChatGPT or showing a warning the user can override (Microsoft Learn, Purview and generative AI apps page, read 12 Nov 2026). Purview endpoint DLP lists a paste-to-supported-browsers activity that evaluates the content being pasted, regardless of how its source item was classified (Microsoft Learn, endpoint DLP page, read 12 Nov 2026).
Those controls cover paste and upload by the user. A page read by an agent is a different path, since the agent collects content from the page it is on, and the user may never paste anything. The ai-dlp-on-the-endpoint guide on this site covers that gap in detail.
A control set a security team can deploy this quarter
Order the work by effort and effect.
- Set extension policy to default deny on managed Chrome and Edge profiles, and keep a reviewed allowlist by extension ID.
- Restrict the permissions an allowed AI extension may hold through ExtensionSettings, and set IncognitoModeAvailability so private windows do not bypass extension controls.
- Inventory what is installed today on a sample of devices before you enforce, so you know which AI extensions people rely on.
- Decide whether agentic browsers are allowed at all. If not, block installation with your application control tooling. If yes, define which accounts may be used in them.
- Turn on browser paste and upload DLP for the AI sites you sanction, and warn or block for the rest, using Purview or your DLP product's browser component.
- Require confirmation for actions that change data or spend money, wherever the agent product offers the setting.
- Log extension installs, removals and permission changes, and alert on a new extension with all-URL host access.
How to test a vendor that says it covers browser AI
Install one AI extension, one agentic browser and one web chat site on a test machine. Ask the vendor's product to show each in its inventory. Paste a test string that matches a sensitive pattern, upload a dummy file and check whether the product records the event and whether it can stop it before it completes. Open the same chat in a second browser and in a private window and check again. A product that sees only the managed browser has said what it covers, even if it did not say so in the data sheet.
Then ask what the product does about an agent that reads page content without any paste. If the answer is a recommendation to restrict sites, that is a policy control, and it is worth having. If the answer is that the product observes the agent's actions, ask to see one.
Where this sits in the index
- Agent Surface MapWhich of six agent surfaces each of 16 tools sees
- Action Control MatrixWhich of six agent actions each tool can stop
- The ranked index16 tools on seven published criteria
Related guides
Sources
- Perplexity: Comet, read 12 Nov 2026
- Google: Chrome extension permission warnings, read 12 Nov 2026
- Microsoft Learn: Edge extension management reference, read 12 Nov 2026
- Google: Chrome Enterprise Help, ExtensionSettings policy, read 12 Nov 2026
- Google: Chrome Enterprise Help, extensions in Incognito, read 12 Nov 2026
- Anthropic: prompt injection defenses in browser use, read 12 Nov 2026
- OWASP GenAI: LLM06 excessive agency, read 12 Nov 2026
- Microsoft Learn: Purview and generative AI apps, read 12 Nov 2026
- Microsoft Learn: endpoint DLP, read 12 Nov 2026
Last reviewed 12 Nov 2026.