What stops an AI agent at a login?
Four walls come up again and again in what people post about their agents.
The agent has no saved login for the site, so it asks you, or it stops. Many people then paste the password into the chat, which puts it in front of the model and its logs.
The site sends a six-digit code by text or email, or asks for an authenticator code. One user put it this way about an agent and an email code: it "won't get it and enter it, you have to do it" [1].
Some sites check whether a browser is run by a person and stop the rest. One OpenAI forum thread from August 2025 has several people reporting the agent's browser looping on a "Verify you are human" check [2]. Banks, Amazon and Delta have all moved to keep agents out [3].
Agent makers block some sites, and some agents stop at signed-in pages. In July 2026, a user on OpenAI's forum posted ChatGPT's own reply that it was "required to stop controlling it once it becomes authenticated" [4]. Agent makers change these rules often, so check your agent's own help page first.
What to do about each wall
Most agents start each task in a fresh browser, so they lose the signed-in session between runs. Saving the session after the first login and restoring it next time fixes most of these. See /guides/keep-ai-agent-signed-in.
They can, if the code is somewhere the agent is allowed to reach. An authenticator key saved in a vault can make the code. An inbox that belongs to the agent can catch emailed codes and sign-in links. Codes sent to your own phone still need you. See /guides/authenticator-codes-ai-agents and /guides/ai-agent-inbox-verification-codes.
No. A password in the chat is in front of the model, in its logs and in anything it writes out. Keep it in a vault that types it into the page for the agent. runnerID does this today: the model never sees the password. See /guides/stop-pasting-passwords-into-prompts.
Then the honest answer is to use what the site offers instead: its own app connection, a staff or delegate seat, or doing that step yourself. Trying to slip past a site's checks breaks its rules and puts your account at risk. runnerID works on your own accounts with your approval, and it does not try to get around blocks.
Use a tool built for logins. runnerID already types the password into the page, makes and catches codes, keeps sessions and asks before risky actions. Coming to runnerID: your agent will be able to hand it the whole stuck step. runnerID does it in its own browser, asks you before anything that sends or pays, and reports back where it stopped.
See plans →Questions
Why can't my AI agent sign in to Google?
Google often treats a new automated browser as an unfamiliar sign-in and stops it, which people report as a "Couldn't sign you in" screen on OpenAI's forum. Where Google offers a connection to your agent's app, use that instead of a browser login.
Do websites block AI agents?
Many do, especially banks, big retailers and airlines, and some block them on purpose. We respect that. Where a site offers a staff or delegate seat or its own connection, use that.
Why does my agent stop after I log in for it?
Some agents are built to stop at signed-in pages, or block certain sites outright, and some sites block automated browsers. That is the agent maker's or the site's rule, not a bug you can fix in settings. Agent makers change this often, so check your agent's current help page.
Is it safe to give my AI agent my password?
Not in the chat. Use a vault that types the password into the page, give the agent its own seat where the site offers one, and require your approval for risky steps. See /guides/ai-agent-own-seat-not-your-password.
Can runnerID solve "Verify you are human" checks?
No. When a site asks for a human check, you do that step yourself.
Related guides
Staff, delegate and service provider seats, and how to take access back.
Read →How to check whether a task finished, and what a useful run log shows.
Read →Saved sessions in Playwright, Browser Use, Browserbase, Kernel, Steel and runnerID.
Read →Sources, checked October 24, 2026
- ›1. Email code quote, a Hacker News comment on OpenAI Dots: https://news.ycombinator.com/item?id=49904198
- ›2. "Verify you are human" loop, OpenAI forum, first post 2025-08-07: https://community.openai.com/t/chatgpt-agent-not-passing-cloudflare-verification-for-chatgpt-com/1336765
- ›3. Banks, Amazon and Delta, Android Headlines, 2026-10-17: https://www.androidheadlines.com/2026/10/banks-retail-websites-block-ai-agents-promises.html
- ›4. "required to stop controlling it once it becomes authenticated", OpenAI forum, 2026-07-15: https://community.openai.com/t/agent-mode-menu-disappeared-after-recent-performance-issues-plus-plan/1386007?page=3
- ›5. Google "Couldn't sign you in", OpenAI forum, first post 2025-07-28: https://community.openai.com/t/openai-agent-cannot-login-to-google/1326900
OpenAI and ChatGPT are trademarks of OpenAI. Google is a trademark of Google. Amazon is a trademark of Amazon. Delta belongs to its owner. Other product names belong to their owners. runnerID and Torchrunner are not affiliated with, endorsed by or sponsored by any of them.