Skip to content

Is It Safe to Let AI Agents Use Your Browser and Logins?

Is it safe to let AI use your browser? What Claude in Chrome, Comet and Gemini can do with your logins, what the vendors advise, and a safer setup.

Janis Zesiger5 min readShare
An AI agent working in a separate cloud session that holds a single login, while the person's own laptop with email, bank and password vault stays out of its reach

Letting AI agents use your browser is safe enough for research and low-stakes tasks, as long as they cannot reach your email, bank or password manager. Claude in Chrome, Comet, Gemini and ChatGPT all have safeguards, but none can rule out being hijacked by a web page.

What can AI reach when you let it use your browser?

A browser agent is an AI that reads a web page, decides what to do, then clicks and types like a person. In your own browser it uses your logins, so to every site the agent is you.

Diagram of the risk when you let AI use your browser: an AI agent's cursor inside a browser window, with lines reaching out to the person's signed-in email, bank, online shop, cloud drive and password vault
In your own browser, the agent inherits every login you have open, not only the one its task needs.

What are the risks of letting AI use your browser?

The main one is prompt injection: orders hidden in a page, an email or an image that the agent mistakes for yours. The UK's National Cyber Security Centre warns it may never be totally mitigated, and OpenAI calls it "unlikely to ever be fully solved". Most published attacks were fixed, but each shows a failure that stays possible.

Risks of AI agents in your browser

Is Claude in Chrome safe? Comet, Gemini and ChatGPT compared

In Anthropic's own tests at the August 2026 launch, 3.8% of attacks that reached Claude Opus 5 succeeded without its extra safeguards; with them, 0% did on Opus 5 and Sonnet 5 and 0.3% on Fable 5. Three weeks earlier, Zenity Labs had shown working attacks on Claude in Chrome, Gemini in Chrome, Comet, OpenAI's Atlas and Microsoft's Copilot in Edge.

Claude in Chrome

All paid Claude plans

Pros

  • Checks each action before it runs
  • Asks before opening financial sites

Cons

  • Screenshots of your tabs become part of the chat

Perplexity Comet

Agent on paid Perplexity plans

Perplexity describes its layered defences. Blocking sensitive sites in comet://settings/assistant is off by default.

Pros

  • Checks every page it reads for hidden orders
  • Confirms before sending mail or ordering

Cons

  • Researchers bypassed one of its security fixes twice

Gemini auto browse in Chrome

US, Google AI plans

Google calls monitoring your tasks the most important protection.

Pros

  • A second AI checks each action against your goal
  • Never sees passwords from Google Password Manager

Cons

  • Works in your own Chrome, with every login there

ChatGPT cloud browser

ChatGPT Work, paid plans

Pros

  • Runs on OpenAI's servers with its own cookies
  • You sign in through a form the AI cannot read

Cons

  • Its sign-ins stay for later tasks until they expire

Which browser should an AI agent use?

The vendors agree: keep the agent out of the browser where your important accounts live. With one shop account, a misled agent can spoil one order; in your everyday browser, it can reach your inbox.

  • Research, price comparisons, public pagesA browser with no sign-ins, such as a disposable cloud session
  • A quick job in a tab, while you watchClaude in Chrome or Gemini, in a separate Chrome profile
  • A recurring job in one or two accountsA saved cloud profile with only those logins, signed in by you
  • The agent works while you are awayA cloud browser you can watch from your phone
  • Company accounts or customer dataAsk your IT team first

What should you never let an AI agent do?

Anthropic, OpenAI and Google warn against all of these. Anthropic also strongly advises against letting its browser agents handle medical information or other people's personal data.

Never let an AI agent

  • Use your main email account
  • Sign in to banking, investments or crypto wallets
  • Work where your password manager has the vault open
  • Get a password, two-factor code or card number through the chat
  • Pay, sign, accept terms or delete data without your final click
  • Run open-ended orders such as "handle my inbox" unattended

How to let an AI agent use your logins safely

Browser.lol is a cloud browser for AI agents: your AI app drives a separate Chrome session on our servers through MCP, a standard way for AI apps to use tools. The agent connection needs Premium; saved profiles are Plus and Premium, currently in beta.

A connected agent can use every session and saved profile on your account and read its Anonymous Mail, so keep only profiles there that it may use.

  1. Connect your agent

    Add the MCP server to Claude, ChatGPT or Cursor and sign in.

  2. Limit its logins

    Start a Chrome session with Persistent session on and sign in to one account.

  3. Sign in yourself

    When a login comes up, click Take over, type the password and code, then Hand back.

  4. Watch, then disconnect

    Keep the live view open; Disconnect agent on the AI agents page ends its access.

The Browser.lol live view after the person took over from the agent: a sign-in page inside the session, with the bar reading You're in control and a Hand back button
While you are in control, the agent's actions wait, but it can still read the page: never click a site's show-password button.

Frequently asked questions

Can an AI agent steal my passwords?

Rarely by reading them: with 1Password for Claude, Anthropic says your passwords and one-time codes never reach Claude. The real risk is a misled agent acting in accounts already signed in, or typing your details into a phishing page.

Does a cloud browser stop prompt injection?

No. Browser.lol does not scan pages for hidden orders: a separate session limits what a misled agent can reach, and what the agent reads still goes to your AI provider.

Who is responsible if an AI agent makes a mistake?

You are. Google, for example, makes you responsible for what Gemini does in Chrome, and the same holds for an agent you run in Browser.lol.

Janis Zesiger

Founder of Browser.lol and Guard.ch

Janis Zesiger founded Browser.lol and Guard.ch and runs Zesiger.net in Switzerland. Janis builds the isolated session platform behind both products and writes about browser security, privacy and AI agents that use the web.

All articles