Your AI Assistant Can Now Sign In Without Seeing the Password. Stop Pasting Passwords Into Chats.
Password managers have started handing AI assistants a login with your approval, and without the password. Here is the office habit to build around it.
By John Martines · October 31, 2026 · 6 min
1Password launched a way for AI assistants to sign in to an account after you approve it, without ever seeing the password. Here is what that changes for a small business, and the rule to write before anyone shares a login with an AI.
The office manager at a dental practice asks an AI assistant to pull last month's statement from the lab's supplier portal. The assistant reaches the login page and stops. "I need the login for the supplier portal." So she does what feels natural. She opens the shared password sheet, copies the password and pastes it into the chat.
The statement arrives a minute later. The password stays behind, in a chat history the practice doesn't control, attached to an account anyone with that history can now open.
That shortcut happens in offices every day. Since last week, there is a better way to do it.
A password manager is an app that stores your logins in an encrypted vault and fills them in for you. 1Password is one of the best known.
On October 22, 2026, 1Password launched 1Password for Agents. An agent is an AI assistant that does things on your behalf, like opening a website, filling in a form or downloading a file. When an agent hits a login page, it asks 1Password for the login. You see which login it wants and the reason it gave, and you approve it in 1Password on your own phone or computer. 1Password then fills in the sign-in form itself, including the one-time code if the account has one. The assistant gets signed in, but it never sees the password and never keeps a copy.
1Password named Meta's Muse, Anthropic's Claude and OpenAI among the AI companies it is building with. Coverage of the launch lists slightly different partner names, so check 1Password's own list for the assistant you use. Reports say it is included with 1Password Individual and Family accounts; the launch did not spell out the business plans.
How it got here:
- October 8, 2025: 1Password and Browserbase, a company that runs web browsers for AI agents, announced agentic autofill, a way for AI agents working in its browsers to use logins stored in 1Password.
- July 16, 2026: 1Password and Anthropic brought approved logins to Claude, so Claude could use a credential without the secret being shown to the model.
- October 22, 2026: 1Password for Agents opened the same approve-then-fill approach to several personal AI assistants at once.
Most small offices run on shared logins: the supplier portal, the bank's bill pay, the shipping account, the state licensing site. Those logins live in a spreadsheet, a sticky note or one person's head. Now AI assistants want to use them too.
Three things change.
A pasted password is a copied password. Once it is in a chat, it sits in that chat's history, maybe on a personal account, maybe on a phone that leaves with an employee. You can't tell who has seen it, and you can't take it back without changing it everywhere.
Approval moves to a person, on purpose. With an approve-then-fill setup, every sign-in by an assistant shows up as a request a named person accepts. That is a record, and a pause, that pasting never gives you.
The approval covers the door, not the room. Once an assistant is signed in, it can do whatever that account can do. If the supplier login can place orders, so can the assistant. The approval tap matters, and so does which account you let it use.
Our earlier piece, [A Backup App Was Used to Reach Into Microsoft 365](/articles/apps-that-hold-your-keys), was about outside tools that hold keys to your accounts. An AI assistant is one more of those tools. This is how you hand it a key you can watch.
1. Make one rule this week: no passwords in chats. Not in ChatGPT, Claude, Copilot, Gemini or a text to a coworker. If a password has already been pasted into a chat, change that password. It takes five minutes and closes the copy you can't see.
2. Get the shared logins out of the spreadsheet. If your office doesn't use a password manager, that is the first move, with or without AI. Put the shared logins in a shared vault, one folder per job (purchasing, shipping, payroll), and give each person only the folders they need. Your IT partner can set it up in an afternoon.
3. Decide which logins an assistant may ask for. Write a short list in three columns:
- *Yes:* read-only or low-risk accounts. Supplier catalogs, shipping tracking, the scheduling system's reports.
- *Yes, with a second look:* accounts that can place orders or send messages. The assistant may sign in, but a person checks the cart or the draft before anything goes out.
- *Never:* bank accounts, payroll, tax and licensing sites, the email admin account, and anything that can change who owns an account.
Keep the "never" logins out of any vault an assistant can reach.
4. Turn on the approval step, and make it go to the right phone. If you use 1Password and an assistant on its partner list, connect the two and approve each sign-in yourself. The person who approves should be the person responsible for that account, not whoever happens to be at the computer.
5. Use the least powerful login that does the job. Many portals let you add a second user with fewer rights, like "view only" or "can order up to a limit." Give the assistant that one. If it is ever misused, the damage stops at the limit.
6. Keep a short habit for odd requests. If an assistant asks for a login that has nothing to do with the job you gave it, say no and look at what it was reading. A web page or a file can carry hidden instructions that steer an assistant off course. The approval prompt, with its stated reason, is where you catch that.
- Does the assistant ever see the password, or does the password manager fill it in?
- Who approves each sign-in, and on which device?
- Can we see a list of every sign-in the assistant made, with dates?
- Does this account have a lower-rights user we could give the assistant instead?
- If we stop using the assistant tomorrow, how do we cut off its access in one step?
Password managers, browsers and the AI companies are all building ways for assistants to sign in on a person's behalf. Expect more of them, expect business plans to follow, and expect the approval prompt to become as normal as the text message code you get today.
You don't need to buy anything new this week. You do need one rule: passwords never go into a chat. Then decide which accounts an assistant may ask for, and who says yes. That way, when your assistant asks for a key, you hand it one you can watch and take back.
---
Sources (checked October 28, 2026). 1Password for Agents, the approval and fill flow, and the companies 1Password is building with: 1Password, press release announcing 1Password for Agents as a trust layer for personal AI agents (Oct 22, 2026). Launch coverage, including the one-time code, the stated reason in the prompt and differing partner lists: 9to5Google (Oct 22, 2026), The Next Web (Oct 2026), iPhone in Canada (Oct 22, 2026) and Let's Data Science (Oct 2026). Availability on Individual and Family accounts: Winbuzzer (Oct 23, 2026); business plans not announced at launch per the same coverage. Approval covering the sign-in, not later actions: SmartScope analysis of the launch (Oct 2026). Browserbase integration: SiliconANGLE, on 1Password's agentic autofill integration with Browserbase (Oct 8, 2025). Claude integration: 1Password and Anthropic press release on secure credential access for Claude (Business Wire, July 16, 2026).