A fake sign-in page is the oldest trick in phishing, and AI agents give it a new disguise. Because an agent connects to your email, calendar, shopping and sometimes payment accounts, a page that looks like "re-connect your agent" or "approve this task" can harvest the keys to all of them at once. The defence starts with knowing how real agents handle sign-in.
The clearest vendor statement comes from Anthropic: Claude Duet's sign-in form "only appears inside the app, during a sign-in you asked for," and "never arrives as a link, a text message, an email, or a website other than claude.ai" (Anthropic). Other vendors have not published an equally explicit rule, but their documented flows point the same way.
How do real agents handle sign-in?
There are two separate sign-ins to think about: signing in to the agent itself, and the agent signing in to other services for you.
| Agent | How the agent signs in to services | What the vendor says about your password |
|---|---|---|
| Claude Duet | Sign-in form inside the Duet app only | "Claude Duet never sees your passwords, browser cookies, sign-in tokens, or other credentials" |
| Meta Muse | Username and password go to secure storage and are injected into the browser when needed | The agent "only ever sees a 'surrogate' token"; email connector filters reset links and magic links |
| Grok Bot | Vault, a connected password manager, or a Secure Form | Signs in "without seeing the values"; SSO, passkeys and CAPTCHAs use a computer takeover |
| OpenAI dots | Saved passwords through a secure form | Not exposed to the model (per launch coverage; OpenAI's FAQ blocked our fetches) |
Sources: Anthropic, Meta, xAI. As of November 2026.
The common thread: real agents keep passwords out of the chat and away from the model, and they collect them inside their own app or a dedicated secure form. A request to type a password into a web page you reached from a message does not fit any of these flows. More detail is in does an AI agent see my passwords? and agent takeover, passwords and 2FA.
What do fake agent sign-in lures look like?
The FTC reports consumers lost $15.9 billion to fraud in 2025. We found no FTC figure specific to AI agents, but agent-themed lures follow familiar patterns:
- Urgent message (email, text, ad)
- Link to lookalike page
- Fake "reconnect your agent" form
- Password and code captured
- Attacker signs in to your real account
Common themes to expect:
- "Your agent needs you to re-authenticate" to finish a task, often citing a booking or purchase in progress.
- "Approve this payment" with a link, when real approvals appear inside the agent's app or the payment provider's app. Claude Duet, for example, shows the exact action and waits for your approval inside the app.
- "Your waitlist invite expires today" with a sign-in link. See fake AI agent apps, extensions and invites.
- "Security alert: unusual agent activity" pushing you to log in to "review" it.
- Ads in search results for the agent's name pointing to a lookalike domain.
Texted versions are covered in depth on our sister site's guide to SMS spoofing and AI agents.
Can a prompt injection produce a fake sign-in request?
It can try. Text in an email or web page your agent reads could tell it to show you a link or ask for a code. This is one reason vendors filter credentials out of what the agent handles: Meta says Muse's email connector "filters out one-time tokens, password reset links, and login magic links." If your agent ever asks you to paste a password, a one-time code or a reset link into chat, treat that as a sign something has gone wrong and deny it. Our page on indirect prompt injection examples explains how hidden instructions reach agents.
How do you check a sign-in request is real?
- Do not clickIgnore the link in the message, however official it looks.
- Open the app yourselfUse the installed app or a bookmark for the vendor's site.
- Look for the request thereA real approval or sign-in will be waiting inside the app.
- Check the addressFor Claude Duet, the only website sign-in is claude.ai; check spelling carefully for every vendor.
- Use a password managerIt will not autofill on a lookalike domain, which is a useful warning.
Passkeys help too. Because a passkey is tied to the real site's domain, it will not work on a lookalike page.
It also helps to know what a real approval looks like for your agent before you need it. Open the app now and find where pending approvals and connected accounts are listed, so a message claiming something is waiting can be checked in seconds.
What if you already entered your password?
- Go to the real site directly and change the password. Change it anywhere else you reused it.
- Sign out all other sessions and review recent sign-ins.
- Turn on two-factor authentication, preferably with an authenticator app or passkey rather than SMS.
- Check connected apps and agent permissions; revoke anything unfamiliar using our revoke agent access guide.
- If you entered card or bank details, call your card issuer or bank immediately.
- Report the page at reportfraud.ftc.gov, and to the FBI's IC3 if you lost money. You can also report it to the vendor through its in-app support.
Bottom line
Real agents ask for sign-ins inside their own apps and keep your passwords away from the model. Anthropic is the only major vendor that states in so many words that its agent's sign-in never arrives as a link, but the habit works for every agent: never sign in from a message, always open the app yourself.



