Skip to content

AI Agent Stuck on a CAPTCHA or Login? Take Over, Hand Back

Your AI agent is stuck at a CAPTCHA, login or 2FA code? Why it should not solve them, what ChatGPT and Claude do, and how to take over and hand back.

Janis Zesiger5 min readShare
An AI agent's cursor waiting beside a sign-in form and a CAPTCHA checkbox in a cloud session, while a person completes the step on a phone

When an AI agent gets stuck on a CAPTCHA or a login, it should not try to get past it alone: it should stop and ask you. You take over the same browser, do that one step and hand control back.

Why does your AI agent stop at logins and CAPTCHAs?

A CAPTCHA is a test that tells people from software, and your agent is software. Many are invisible: Cloudflare Turnstile shows a box only when something looks risky, and reCAPTCHA v3 scores every visitor without showing anything.

Where browser agents stop, and who should take the step
CheckpointWhy the agent stopsWho does it
CAPTCHA or "Verify you are human"Built to block automationYou
Password sign-inYour credential, not the agent'sYou
Two-factor (2FA) code or push approvalIt arrives on your phoneYou
Payment, terms, new accountIt commits you to somethingYou decide

Can an AI agent solve a CAPTCHA?

Older visible puzzles, often yes. New puzzle types and checks that score the browser, mostly no.

40%

Best agent on puzzles, 2025 models (people 93.3%)

90%

Newer models on one puzzle type, 2026, at best

5.9%

Best model on 27 new puzzle types (people 98.8%)

23%

Paid solving services on reCAPTCHA v3, average

Sources: Open CaptchaWorld, its 2026 follow-up and Broken Gates, where on reCAPTCHA every fresh browser scored like a bot and agents passed only in an established Chrome profile signed in to Google. Bypassing is no shortcut: many sites forbid it, and the banned account is often yours.

What ChatGPT, Claude, Gemini and Browser Use do at a CAPTCHA

How browser agents handle logins and CAPTCHAs, as of November 2026
AgentAt a loginAt a CAPTCHAWhere you step in
ChatGPT (cloud browser)Secure form the model cannot seeAsks you to take overA link, phone or computer
Claude in ChromePauses, or 1Password fills itPauses; will not bypassYour own Chrome
Gemini in ChromeYou, or Password ManagerPauses for youThe task tab in Chrome
Browser Use CloudSaved logins or live viewSolves some itselfLive view link
Browser.lolAsks youAsks youLive view, any device

ChatGPT's old agent mode is gone: browser tasks now run in the cloud browser of ChatGPT Work, not on the Free and Go plans. Claude in Chrome works beside every account you are signed in to, so a misled agent can reach them all, which is the main safety risk.

How to take over and hand back in Browser.lol

Your agent drives a Chrome session in the cloud, and you watch it in the live view (the session, live in a page). The agent connection needs Premium.

  1. Connect your agent

    Add Browser.lol to Claude, ChatGPT or Cursor once, with the setup guide.

  2. Open the live view

    Your agent sends its link when the session starts.

  3. Do the step

    At "Your agent needs you", sign in, type the code or tick the box.

  4. Hand back

    Click the "Done, hand back" button. If nothing happens, reply "Done, carry on".

The Browser.lol live view of an agent session showing the request Your agent needs you, the agent's message asking the person to sign in, and the Done, hand back button at the bottom of a sign-in page

You can take over without a request too; the agent's actions wait until you hand back. While you are in control the agent can still read the page: a password field shows only dots, but a code and every page after sign-in are readable. Never type a secret into the chat.

Stay signed in so your agent rarely has to ask

A saved profile (Plus and Premium, currently in beta) keeps cookies, logins and history between sessions. It works only with the browser it was made with, such as Chrome, and an unused one is deleted after 30 days unless you pin it; more in a cloud browser that stays logged in.

Set up a signed-in profile for your agent

  • Ask your agent to start a session saved as a new profile, such as "Shop"
  • Sign in yourself in the live view, including the 2FA code
  • End the session: the profile is saved when it ends, not while it runs
  • Next time, tell the agent to use "my saved profile Shop"
  • Never copy cookies from your own browser into it: a stolen session cookie gets past 2FA

When is a cloud browser the wrong tool for your agent?

Browser.lol never solves or bypasses CAPTCHAs and does not make an agent undetectable, so sites can still challenge or block a session. Agents drive Chrome and other Chromium-based browsers, not Firefox or Tor Browser.

  • The site offers an official API or an app for agentsUse that: no CAPTCHA, no hand-off
  • The work is in tabs you already have open, signed in as youClaude in Chrome or Gemini in Chrome
  • You want to step in from your phone and keep the agent out of your own browserA cloud session with a live view, such as Browser.lol

Frequently asked questions

How do I tell my agent to stop and ask instead of guessing?

Write clear stop rules into its custom instructions: never solve a CAPTCHA, never type a password, 2FA code or card number, and ask me first. After you hand back, it should look at the page again before it continues.

Can my agent read a confirmation code by itself?

Yes, for accounts it creates with Anonymous Mail, a throw-away inbox in Browser.lol: the agent reads the code itself, with no hand-off.

Can someone else open my agent's live view?

Only if you let them: Browser.lol's live view needs your sign-in or a share link you create. Browser Use's live view link lets anyone holding it click and type.

Is it legal for an AI agent to log in for me?

Signing in to your own account is generally your decision, but each site's terms apply and some forbid automated access.

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