OnlyRead
Star102

Is the browser the AI agent attack surface everyone misses?

A round white bot with two black eyes sits on a cloud beside a desk with a closed laptop, an empty mug and a clock. It reads a browser feed marked with green checks, while its pink cursor is stopped by a pane of glass sealed with a pink padlock in front of the send and like buttons.

The short answer

Yes. Most agent security today guards the tools an agent calls, through MCP servers, connectors and the gateways in front of them. Agents like Grok Bot also open a browser and use any website the way you do, and nothing outside the agent checks what they click there. That browser is a new attack surface, and OnlyRead guards it.

Picture the end of a workday. You close your laptop, and your Grok Bot carries on without you. That’s by design: xAI says work “continues while your laptop is closed”, and your primary Bot “checks in with you on its own”.12 The security researchers at NewCore looked at what that means in a blog post this week, and they found Grok Bots reaching into connected apps on their own while their owners were away.3 Their post is about the tools an agent calls, and so is almost every conversation about agent security today. We want to take it a step further, because there’s a second surface that hardly anyone is looking at yet, and it’s the browser.

What does a Grok Bot do while you’re away?

Quite a lot, and xAI is open about it. Its docs say your Bot “doesn’t only wait for instructions. It notices work it can take off your plate and offers to handle it”, and it lets you know when something needs your attention.4 Routines you set up run on schedule whether you’re around or not, and when background work raises an approval card, the card expires after 30 minutes if nobody answers it.5

So while you’re away, the Bot’s own judgment decides what happens next. xAI says the Bot “offers first” and “asks before risky actions”,4 which is a sensible default, and it’s the Bot that judges which actions count as risky. Muting its notifications doesn’t pause any of this, because they “control alerts, not the check-ins themselves.”4 That’s exactly what you got the Bot for. It also means that whatever guards your accounts has to work with nobody in the room.

Security has caught up with the tools

When people talk about securing agents today, they mostly mean the tool layer. An agent gets tools from MCP servers or an app’s own connector, each tool calls the app with your access, and a gateway in the middle can log every call and hold the ones that shouldn’t go through. A Grok Bot uses this layer for apps like Gmail, Slack or Google Calendar, whose native connectors hand it tools that call the app directly.1

NewCore’s research shows what this layer can do. Its team ran two Grok Bots through its own approval gateway for five days and counted 31 proactive requests to connected apps. All were reads, and the gateway held each one before it reached any data. Since then, xAI has given team admins a Proactive messages switch that turns check-ins off. It’s on by default.3 That’s good work, and if your company runs a gateway, keep it. But look at where those 31 requests went. Every one passed through a connector, and the gateway saw them because they were tool calls.

Why is the browser a new attack surface?

A connector only exists when an app offers one, and many of the apps people care about most don’t. This is where Grok Bot’s design gets interesting. It doesn’t borrow your Chrome. It gets a cloud computer of its own, which xAI describes as “separate from the Mac or Windows computer in front of you”, with its own browser. You sign in there to the accounts you want the Bot to use, its “browser sessions persist between tasks”, and every Bot on your account shares the same sign-ins.1 Your browser stays yours, the Bot’s stays its own, and the Bot can reach any service, with an app or without one. xAI’s advice is to “use the browser for services without an app”.1 That’s how the Bot works on LinkedIn or WhatsApp Web, which offer no way to connect an AI assistant with read-only access. We covered why LinkedIn has no such option in an earlier post.

A click in that browser isn’t a tool call, so a gateway never sees it. Admin controls split along the same line. If your admin blocks an app, xAI notes that this “doesn’t block that service’s website,” and closing off the website as well takes Network Controls, which only the Enterprise plan has.1 What’s left is the Bot’s own Auto Review, which decides which actions need your approval. It’s useful, and xAI is honest about its place: it “is model-based and should complement, not replace, least privilege and explicit approval boundaries.”5 In the browser, that boundary doesn’t exist yet. As agents get better at using websites the way people do, more of their work will happen there, and it needs a guard of its own.

PathAppsWhat can stop a send
Native connectorGmailSlackGoogle CalendarNotionThe app’s tool toggles, Grok Bot’s approvals, or an approval gateway
The Bot’s browserLinkedInWhatsApp WebInstagramGrok Bot’s approvals, or OnlyRead in the Bot’s browser

A guard for the browser

That guard is what we built OnlyRead to be, and as far as we know it’s the only tool today that covers this surface. OnlyRead sits on the boundary between the agent’s browser and the internet, inside the Bot’s own Chrome. From there it sees what the agent reads on each website and every action it tries to take. It works independently of the Bot, so it makes no difference whether a click started in your chat, in a routine or in a check-in at 4 in the morning, or whether Auto Review already let it through. What you do with that view depends on your plan:

  • Free and Pro keep it simple. For each site, you choose whether the agent’s browser is read-only there. On a guarded site, the Bot can still open your LinkedIn messages, read a profile and run a search, but a message, a post, a like or a connection request is stopped before LinkedIn’s server ever sees it.
  • Enterprise adds a full audit and finer control. You can see which sites each browser agent visited and which actions it took, and set rules per site and per action: let it through, block it completely, or hold it until a person approves.
  1. YouChoose the sites the Bot’s browser keeps read-only
  2. Grok BotWorks in its browser while you’re away, and its Auto Review decides what to ask you
  3. OnlyReadRefuses requests that would change something
OnlyRead covers the sites you guard, in the Bot’s browser. Connected apps go around it.

When you set it up with the bot setup link below, OnlyRead guards the Bot’s browser all the time, not only when it notices an agent at work. The Bot can’t loosen those settings either. The browser is managed from your OnlyRead account, and the sign-in the link gives the Bot can add sites to your list but never remove one. That’s what lets you close your laptop without wondering what the Bot might send.

  1. You close your laptop for the night. The Bot keeps working on its cloud computer.

  2. A routine opens LinkedIn in the Bot’s browser and reads your new messages.

    The read goes through and is logged
  3. A check-in decides to answer one of them. Auto Review lets it go.

    OnlyRead refuses the send
  4. You’re back. Nothing went out on LinkedIn that you didn’t write.

An example night. OnlyRead checks what the Bot does in its browser, not in its connected apps.

Set it up on a Grok Bot

  1. Open Set up OnlyRead on your bot and sign in. Setting up from a bot needs OnlyRead Pro, which keeps your guarded sites in your account.
  2. Choose Grok Bot and paste the message OnlyRead copies for you into the Bot’s chat. It gives the Bot your permission to add OnlyRead from the Chrome Web Store to its own browser and connect it with a one-time link.
  3. Choose the sites to guard. When the Bot says the page is done, its browser shows up under Browsers on your account page, marked as Grok Bot. Switch on the sites it should only read, such as LinkedIn, X and WhatsApp. Until a site is on your list, the Bot’s browser doesn’t guard it.

The general guide explains how read-only browser protection works, including why a search goes through when a send that looks alike doesn’t.

Test it on an account you own

  1. Use something you own, such as a private test account on X, where nothing reaches anyone else. Sign in to it in the Bot’s browser, guard that site on your account page, and don’t connect the X app in Grok Bot for this test.
  2. Ask the Bot to read, for example: “In your browser, open my latest post on X and tell me what it says.” The answer should come back as usual.
  3. Ask for the small change: “In your browser, like that post.” If Auto Review shows an approval card, allow it, since that’s Grok Bot’s own check. A click that gets past it is what OnlyRead is for.
  4. Ask the Bot what happened. OnlyRead puts a notice on the page naming the refused change, and the Bot can see it on screen.
  5. Open the post from your own browser and check that it has no like.

What this doesn’t cover

Neither Grok Bot’s settings nor OnlyRead is a guarantee, and OnlyRead has limits worth knowing before you rely on it:

  • Connected apps go around it. What the Bot does through a connected app never passes through its browser, so OnlyRead doesn’t see it. That surface needs its own guard, such as an MCP gateway.
  • Requests its built-in rules can’t settle go to an AI model, which can misjudge one. It runs apart from the Bot and answers one question, whether a request changes data. When it’s unsure, or can’t be reached, OnlyRead refuses, so a doubtful call costs a retry rather than a sent message.
  • All your Bots share one cloud computer and its browser sign-ins, so a guarded site is read-only for every Bot, and for you when you take over the computer.
  • Other routes on the Bot’s computer go around it. OnlyRead checks the Chrome it’s installed in, not commands, scripts or another browser the Bot runs on its computer.

Guard both surfaces

The tool layer now has gateways and people measuring what passes through them, and that’s real progress. The browser deserves the same attention, before it becomes the way most agent work gets done. So our recommendation is simple. If your agents work on their own, guard both layers:

  • The tool layer, with a tool that guards the calls your agents make through MCP servers and connectors, such as a gateway.
  • The browser layer, with a tool that sits between your agents’ browser and the websites they use. That’s OnlyRead.

Questions

Isn’t an MCP gateway enough?

Not for the browser. A gateway sees the calls an agent makes through MCP servers and connectors. When the Bot opens LinkedIn in its browser and clicks Send, no call passes through the gateway, so it never sees it. Keep the gateway for connected apps, and add a check in the browser.

Does OnlyRead see what the Bot does in connected apps?

No. Connected apps don’t pass through the Bot’s browser. OnlyRead covers the other path, the sites the Bot uses in its browser, and there it refuses sends whether you were in the chat or away.

Isn’t turning off check-ins enough?

It stops the check-ins, but not routines you set up or a task you gave the Bot that it reads differently than you meant. And check-ins are on until support, or a team admin, turns them off.

Can I still let the Bot post on one site?

Yes. Leave that site off your guarded list. Every other guarded site stays read-only for the Bot.

Does it work with other cloud agents?

The bot setup supports Grok Bot today, and more bots are coming. Agents that work in your own Chrome, such as Claude in Chrome, use the extension from the Chrome Web Store.

Sources

  1. Use the computer and apps, Grok Bot docs, xAI, retrieved Nov 17, 2026.
  2. Create and manage Bots, Grok Bot docs, xAI, retrieved Nov 17, 2026.
  3. Do You Know What Your Grok Bot Did Last Night?, Ido Hoorvitch, NewCore, Nov 17, 2026.
  4. When your Bot takes initiative, Grok Bot docs, xAI, updated Nov 11, 2026.
  5. Approvals, security, and privacy, Grok Bot docs, xAI, retrieved Nov 17, 2026.