Can an AI agent sign in without seeing your password?
A password should never travel through a chat log or a language model. The safer pattern is a form that types it straight into the sign-in page, on that site alone. GROX follows that pattern on GROX's computer: a secure form in the chat, nothing for the model to read, and a choice to save encrypted or forget within thirty minutes.
Why should a password never pass through a chat or a model?
Chat is a log. Anything you type there can be stored, summarised, or shown in a later thread. A model that sees a password can also echo it, mix it with other context, or treat it as ordinary text. That is the wrong place for a secret.
The same rule applies to API keys. They are not conversation. They belong on the page or in the process that will use them, not in a message the assistant can quote. If a simpler password manager in your browser already fills the field on a site you open yourself, that is often enough; an agent is only useful when the work must happen on a remote computer you are watching.
How does a secure form type a password only on its own site?
The computer reaches a sign-in page and stops. You are not asked to paste the password into the thread. A dedicated form appears, bound to that page. The characters go into the field on the site, not into the model and not into chat history.
If the site later refuses the login, the agent should not keep retrying the same secret. One failed attempt is a signal to stop and ask you. One-time codes and payment pages are the same class of halt: the work pauses until you finish the sensitive step.
| Path | What happens | Who can read it |
|---|---|---|
| Typed in chat | Stored with the thread | The model and later sessions |
| Pasted into a prompt | Treated as ordinary text | The model writing the reply |
| Secure form on the page | Typed only on that site | The site, not the chat |
| Encrypted save for next time | Used later after you agree | Not shown in the thread |
How does GROX handle a sign-in on GROX's computer?
When a site on GROX's computer asks you to sign in, the chat shows a secure form. The password goes straight into that page, never into the chat and never to the AI, and is typed only on the site it is for. Tick save it for next time and it is kept encrypted so GROX signs in by itself later; otherwise it is forgotten within 30 minutes.
An API key goes in the same way and reaches GROX's terminal without being shown. Before GROX's computer signs in with a login you saved, it asks whether to use the saved login, with one tap; you can turn that asking off in chat. It tries a saved password only once. If the site refuses it, it stops and asks you. When the computer reaches a password, a one-time code or a payment page, it can send your phone a link that opens only that screen for fifteen minutes; Done hands the work back. The link is not a sign-in to your account.
- Secure form
- A chat card that sends the password into the page, not into the model.
- Encrypted save
- Optional keep for later sign-ins; GROX still does not see the secret in chat.
- Thirty-minute forget
- If you do not tick save, the password is dropped after the task window.
- API key box
- The same path: the key reaches the terminal without being shown in the thread.
When is a simpler tool the better choice?
If you only need to log into one familiar site on your own laptop, the browser's password manager is simpler. An agent computer is for work you want done in a visible remote session: browsing, filling fields, running commands, then handing files back.
Do not give an agent a password in prose, even as a joke. Do not ask it to remember a password as a chat fact. If GROX learns something about you it says so under the answer, and forget that removes it — but passwords are not that kind of memory; they stay on the form path.
- Keep secrets off the transcript.
- Prefer a form bound to one site.
- Save encrypted only if you want later unattended sign-in.
- Stop after one refused password.
Common questions
Can an AI agent sign in without seeing my password?
Yes, if the password never enters the chat or the model. A secure form types it only into the site's own field. GROX does this on GROX's computer: the secret goes straight into the page, is forgotten within 30 minutes unless you tick save, and a saved login is tried only once.
What happens if I tick save it for next time?
The password is kept encrypted so later sign-ins can happen without typing it again. Before a saved login is used, you are asked with one tap unless you turned that prompt off in chat. If the site refuses the password, the attempt stops and you are asked rather than looping.
Does the same idea apply to API keys?
Yes. An API key should not be pasted into a conversation. It goes through the same secure box, reaches the terminal without being shown, and is never treated as chat text the model can quote back.
Should I still use a normal password manager?
For sites you open yourself, yes. An agent session is extra machinery. Use it when the work must run on a remote computer you can watch, not as a replacement for filling a field on your own browser.
Read how GROX's computer handles sign-in on grox.life and the Help Centre.