How to choose
How should a browser agent handle your logins?
Saved profile, vault fill, 1Password approval or OAuth? Where each browser agent login leaks, which sites forbid agents, and which one to pick.

When your agent hits a login page, the real choice is who holds the secret: you, a vault, the cloud browser, or the site itself through OAuth. For most of your own accounts, I’d log in once by hand into a saved, per-site browser profile and keep the password away from the agent entirely. For banking, airline loyalty and eBay buying, I wouldn’t hand the agent the login at all.
I compared the login docs of eight cloud browsers, 1Password’s agent docs, the consumer agents from OpenAI, Anthropic and Google, and the terms of sites that now restrict agents. If you’re still choosing the browser itself, start with picking a cloud browser for AI agents; this page is about what happens at the sign-in box.
Short on time?
Short on time? Here's my take.
- Your own accounts (default)
- Log in once into a Browserbase Context
- Sign in yourself through live view into one Context per site and user, encrypted at rest, then reuse it read-only. Browser Use or Steel profiles work the same way.
- The agent must re-login on its own
- Steel Credentials API
- Bound to the exact origin, fields blurred, 2FA codes generated server-side, and free during the beta (verified 2026-11-17). Kernel or Anchor managed auth if you can't accept beta.
- A person should OK each login
- 1Password for Agents on Browserbase
- Approval happens on your device and the agent never holds the values. Add your own per-task rule: one grant lasts 30 days.
- The site offers OAuth, an API or MCP
- OAuth, not a browser login
- Scoped, expiring and revocable. The MCP spec builds on OAuth 2.1 with PKCE required.
- Banking, airline loyalty, eBay buying
- Don't hand over the login
- eBay and United restrict agents in writing, and Bank of America says you're responsible for what an agent does with your login.
- No code
- ChatGPT's sign-in form or Claude in Chrome
- ChatGPT's form keeps credentials from the model; Claude in Chrome belongs in a separate profile, paired with 1Password for Claude.
Prices verified . Sponsors cannot buy a verdict or a position.
Here’s the same advice laid out by the kind of account you’re connecting.
| Account | Pattern I’d use | Why |
|---|---|---|
| Your own SaaS dashboards, admin panels, supplier portals | Log in yourself once into a saved per-site profile | No password stored anywhere; the session expires on the site’s schedule |
| Accounts the agent must re-enter unattended | Vault fill at type time (Steel Credentials API) | The model gets names, not values, and the fill is bound to one origin |
| Shared work accounts where a person should sign off | Password-manager fill you approve (1Password for Agents) | Approval on your own device; nothing stored in the agent |
| Google, or any site with OAuth, an API or an MCP server | OAuth | Google may block automated browser sign-ins; tokens are scoped and revocable |
| Banking, airline loyalty, eBay buying | Keep the agent out | eBay and United restrict agents in writing; Bank of America puts the liability on you |
| Personal no-code tasks | ChatGPT’s secure sign-in form or Claude in Chrome | You or 1Password enter the password; the model doesn’t see it, per both vendors |
Seven ways an agent can get signed in #
Every login setup I found is one of seven patterns, and they differ on two questions: who holds the secret, and what can read it on the way to the form. The vendors’ marketing explains the mechanism; their own docs, read closely, also admit where each one leaks.
| Pattern | Who holds the secret | What the model sees | Where it leaks | How you revoke it |
|---|---|---|---|---|
| 1. You log in once into a saved profile | The cloud browser holds your session, not your password | The signed-in pages | Anyone who copies the profile can act as you | Sign out on the site; deleting a Browser Use profile doesn’t |
| 2. Sync cookies from your own Chrome | Your live sessions, copied to the vendor | The signed-in pages | Copies sessions off your device; unscoped; device-bound sessions can break it | Same as pattern 1 |
| 3. Secret typed in at fill time | The vendor’s vault or your worker | Secret names or placeholders, per the vendors | Screenshots, model-written code, live views, page scripts | Rotate the password; delete it from the vault |
| 4. Password-manager fill you approve | 1Password; the agent “never holds credential values” | Nothing, per 1Password | One approval can cover many fills for 30 days | Deny or revoke the grant in 1Password |
| 5. Managed re-login | The vendor stores the password, often with the 2FA seed | Not added to model context, per Kernel | The vendor holds both factors; page scripts can read filled values | Delete the stored identity; rotate password and 2FA seed |
| 6. OAuth or MCP | The site issues a scoped token | A token, not your password | Misuse inside the granted scope | Revoke the grant; tokens expire |
| 7. Consumer secure sign-in form | You type into a form the model can’t see | Not the credentials, per OpenAI | The session persists until it expires | Clear the site’s data in cloud browser settings, which signs you out |
Verified November 17, 2026. Each row’s sources are linked in the sections below.
Patterns 1 and 2 store a session, not a password. You sign in once, by hand, and the browser profile keeps the cookies and local storage; Browser Use’s profile sync copies them from your local Chrome in an explicit CLI step and says plainly that cookie sync “does not import a password manager”. Manus works the same way, capturing cookies and local storage, encrypting them and injecting them into new sandboxes, as Logto describes it.
WorkOS points out what cookie sync loses: the copied session isn’t scoped to the task, and the site’s logs can’t tell the agent’s actions from yours. Logto lists replay of stolen session files and fingerprint mismatches as the risks. Both companies sell OAuth tooling, so read them as advocates, but the risks they list match what the cloud browsers’ own docs say.
Patterns 3, 4 and 5 store a password somewhere and fill it at the last moment, so the model never reads it. The difference is who decides when the fill happens: your code, a person tapping “approve”, or the vendor’s re-login service.
Pattern 6 is the one I’d pick whenever the site offers it. The MCP authorization spec builds on OAuth 2.1, requires PKCE with S256, requires the client to name the resource it wants, forbids servers from passing a token through, and says short-lived tokens should be used. That gives you scopes, expiry and revocation that no saved cookie offers.
At least one security-minded reader on Hacker News agrees. On a 2025 Show HN for a delegated-login tool, user bobbiechen called the approach “account takeover as a service” and argued for OAuth-style scoped consent (Hacker News).
A saved session is a password, so store it like one #
A logged-in profile is worth as much to an attacker as the password that created it. Playwright’s docs warn that saved browser state “may contain sensitive cookies and headers that could be used to impersonate you”, and tell you to keep it out of git.
MITRE ATT&CK technique T1539 says the same thing from the attacker’s side: a stolen session cookie lets someone into a web app as an authenticated user without the credentials, and can get around some multi-factor checks. So a profile that sits in a cloud browser for months is a standing key to that account.
How long do vendors keep your session? #
Retention varies more than I expected, and half the vendors don’t say.
| Vendor | What it keeps | How long |
|---|---|---|
| Browserbase | Context: the Chromium user data directory (cookies, localStorage, IndexedDB, session storage, service workers, autofill), not the HTTP cache | “live indefinitely” until you delete them; “uniquely encrypted at rest” |
| Steel | Profile snapshot of cookies, storage and extensions, up to 300 MB | Deleted after 30 days unused, per Steel’s article |
| Cloudflare | Tabs, cookies and login state while the session is alive | Idle up to 10 minutes; “expiry or a crash loses” the login state |
| Browser Use | “cookies, local storage, and login state” | Not stated |
| Hyperbrowser | Cookies, local storage, session storage and cache | Not stated |
| Anchor | Profiles without credentials; managed identities with credentials and session state | Not stated |
Verified November 17, 2026.
Indefinite storage isn’t bad in itself; it means the cleanup is your job. I’d put a calendar reminder on every Context that holds a sensitive login, and delete it when the project ends.
One profile per site, per user #
Three vendors give the same rule. Browserbase recommends one Context per site or login and warns that concurrent sessions on one Context can make a site force a logout (Contexts docs). Browser Use says “Use one profile per end user” (authentication guide), and Cloudflare says “Use separate browsers when users need separate login sessions” (browser tool docs).
Browserbase also gives you a useful switch: persist: true saves changes when the session closes, and persist: false reuses the Context read-only. My reading of that: log in with persist: true once, then run routine tasks read-only so a confused run can’t overwrite a good session.
Sessions expire, and deleting a profile doesn’t sign you out #
A saved profile doesn’t refresh itself. Browserless puts it well: “a profile is a snapshot, not a live sync” (Browserless). Browserbase tells you to build logged-out checks into your flow so the agent notices when it’s been signed out instead of failing quietly.
The part I’d underline: Browser Use’s sync docs say deleting a profile does not sign out a running browser or revoke the site account. To kill a leaked session you have to sign out on the site itself.
Device-bound sessions can break cookie sync #
The same Browser Use page warns that sites may “bind sessions to a device or network”. Google already does. Its Device Bound Session Credentials tie a session to a hardware key that can’t be exported, and it’s an open W3C standard other sites can adopt.
Google made it generally available in Chrome on Windows for Workspace and personal Google accounts, rolling out from May 25, 2026, on by default with “no administrator control to disable it” (Workspace Updates). Google’s DBSC post says macOS comes in “an upcoming Chrome release”. So a Google session synced from a Windows Chrome can stop working in a cloud browser; that’s the design doing its job, and the fix is OAuth, not a workaround.
If you run browsers yourself instead of renting them, all of this still applies to the profile folder on your own disk.

How do you keep the password away from the model? #
Type the secret in at the last moment, outside the model, and make sure nothing reads it back. Four vendors document a version of this, and their docs differ most on what they admit can still see the value.
Steel’s Credentials API is the most locked-down default I read. Steel’s docs say autoSubmit, blurFields and exactOrigin are all on by default, so the fill is bound to the exact site, the fields are blurred and the form submits itself. A stored totpSecret generates 2FA codes that are “never exposed to the page”, and each credential gets its own AES-256-GCM key, wrapped by an organization key and bound to the org ID and origin.
The catches are in the same doc. It’s “currently in beta”, storage and use are free during the beta (verified 2026-11-17), and credentials are stored “globally against your organization”, so I’d treat the Steel org as your security boundary. Steel’s own article says the feature is Cloud-only and sums up the point in five words: “Stop pasting passwords into prompts.”
Anchor’s secretValues promise “Never logged”, “Never sent to AI”, “Type-time replacement” and “No storage”, with secrets scoped to domain patterns and the * wildcard to be used sparingly (Anchor). It works with Browser Use and with the OpenAI, Anthropic and Gemini computer-use agents.
Skyvern gives the model placeholders and types the real value at input time, but says typed values can appear in later screenshots unless “Mask Secrets” is on (Skyvern). It also tells you to treat Bitwarden API keys like master passwords, which applies to any vault key you hand a cloud service.
Browser Use’s sensitive_data is the most honest page I read. “The model is told only the secret names”, but the values live in the worker, so code the model writes can read them; “Screenshots are not redacted”; a script can read a field back; live views show typed values; and “This is not a sandbox” (sensitive data docs). Its cloud version scopes secrets with secretBindings.allowedDomains to the focused field’s frame only (authentication guide).
Kernel adds a warning that applies to all of them: page scripts, extensions or agents with unrestricted access can read a filled value, so use “the narrowest browser permissions” (Kernel). Browserbase’s version of the idea is templating that “hoists sensitive data as parameters, to be injected at runtime” (Browserbase).
Live-view links are keys too #
Three vendors say plainly that a live-view link is as powerful as the login. Steel says its debug URLs are “intentionally unauthenticated” and should be treated as secrets (Steel’s article). Cloudflare is blunter: “Anyone with a link can control the browser, including the sites it is logged in to” (Cloudflare).
Browser Use says the same of its live_view URL (docs). Cloudflare’s Live View links expire after 5 minutes by default, 1 hour at most (human in the loop); I’d keep every live-view link that short and never paste one into a chat or ticket.
Proxy passwords are the same problem in a different place; I cover that in keeping proxy credentials away from the model.
What does human approval actually approve? #
Less than the phrase suggests: in 1Password’s design, one approval can cover many fills for up to 30 days. That’s still the best option I found for accounts where a person should stay in the loop, as long as you know what you’re approving.
1Password for Agents shows you which agent is asking, then fills the username, password and one-time code itself; “The agent never sees the credentials.” 1Password’s launch post from October 2025 describes human approval by default and an end-to-end encrypted channel built on the Noise framework, and puts the principle in one line: “Raw credentials should never enter the LLM context.”
Browserbase shipped it on October 22, 2026: you approve or deny on your own device, 1Password fills and submits, and sessions can run with zero data retention and no recording (changelog).
Read the grant, not the button #
The fine print is in 1Password’s partner docs, labelled “Partner preview”. A grant lasts 30 days at launch, access tokens last 15 minutes, and “one approval can authorize multiple fills”. Approving each session is something the partner platform has to enforce.
Browser Use’s comparison post says Browserbase’s flow asks for Touch ID or a master password before each fill (Browser Use). That conflicts with the partner docs, and I trust 1Password’s own description of its own product. So if you want a person to approve every task, write that rule into your app; don’t assume the approval prompt does it for you.
Not every 1Password setup is an approval setup #
Two documented setups look similar and work very differently. Browserbase’s 1Password quickstart uses a service-account token: your script pulls the credentials and signs in, so the values pass through your code, and the token “will only be shown once”.
Browser Use’s 1Password guide lists three modes. With a service account on a dedicated vault, “The agent receives the value”, every supported field in that vault “is made available to the run”, up to 20 per run, and it isn’t available with zero data retention. With a personal 1Password account you approve in the app and the agent never sees the password.
One more detail from that page: a fill_submitted result means “the form was submitted, not that the sign-in worked”. Keep your logged-out check either way.
My take: use the personal approval mode, or 1Password for Agents on Browserbase, for accounts a person should sign off on. If you do use a service account, give it a vault that holds only what this agent needs.
Can an agent handle 2FA and passkeys? #
It can type 2FA codes if you give it the authenticator seed, but it can’t use a passkey on its own, and for sensitive accounts I’d keep a person on the second factor.
Five of the vendors I read let you store an authenticator seed so the agent can type its own 2FA codes: Steel, Browser Use, Kernel, Anchor and Skyvern. Anchor also lists email codes and 1Password, and Skyvern adds email and SMS.
Here’s what none of those docs says outright. If the agent holds the password and the seed, it holds both factors, and 2FA stops proving a person was there. WorkOS’s blog makes the same point. For a low-stakes internal tool that may be fine; for anything with money or customer data, I’d keep the second factor with a person.
Passkeys need you, by design #
A passkey’s user check “can only be exercised by the user with a biometric or PIN”, in FIDO’s words. Skyvern lists passkey 2FA as Enterprise-only and says passkey credentials can’t use saved browser sessions (Skyvern).
Standards are coming, not here. FIDO announced an Agentic Authentication working group on April 28, 2026, chaired by people from CVS Health, Google and OpenAI, next to a payments working group that starts from Google’s AP2 and Mastercard’s Verifiable Intent (FIDO). It named no spec dates.
Handing the browser to a person #
Cloudflare frames the human handoff as the answer for logins with MFA, SSO or CAPTCHA that “cannot be bypassed programmatically”: a person signs in through Live View, and a structured handoff can wait up to 30 minutes (Cloudflare). Browserbase’s recommended flow includes logging in once by hand through its live view (Contexts), and Kernel’s managed auth “requests user input when needed” (Kernel).
BrowserAct’s guide is a product pitch, but its rule is right: “Agents should not try to bypass 2FA.” My pick: a person handles 2FA for every sensitive account, and the agent gets the signed-in session afterwards.
Where does each cloud browser stand? #
Only Browserbase and Steel say how they encrypt what they store, only three vendors state how long sessions last, and Cloudflare stores no credentials at all. Here’s what each vendor documents for logins, and where the docs are silent. “Not stated” means I didn’t find it on the pages I read, not that the feature is missing.
| Vendor | What persists | Retention | Credential feature | 2FA | Human takeover | Encryption stated |
|---|---|---|---|---|---|---|
| Browserbase | Context (full user data dir) | Indefinite until deleted | 1Password for Agents; 1Password service account | One-time codes via 1Password | Live view | Contexts encrypted at rest |
| Browser Use | Profile; sync from local Chrome | Not stated | sensitive_data; 1Password |
TOTP seed | live_view URL |
Not stated |
| Steel | Profile up to 300 MB | 30 days unused | Credentials API (beta, free during beta) | Server-side TOTP | Debug URL (unauthenticated) | Per-credential AES-256-GCM |
| Hyperbrowser | Profile; changes off by default | Not stated | Not stated | Not stated | Not stated | Not stated |
| Anchor | Profile; managed identity | Not stated | secretValues; managed re-auth |
Email code, TOTP, 1Password | User signs in (manual auth) | Not stated |
| Kernel | Reusable profile from managed auth | Not stated; health checks and auto re-auth | Managed Auth | TOTP | Asks the user when needed | Not stated |
| Skyvern | Saved sessions | Not stated | Skyvern, Bitwarden, 1Password or Azure vault | Authenticator, email, SMS; passkey (Enterprise) | Not stated | Not stated |
| Cloudflare Browser Run | Session state while alive | Lost on expiry or crash | None; human handoff | Person signs in | Live View, handoff up to 30 min | Not stated |
Verified November 17, 2026. Steel’s retention and profile size come from a Steel article, not its main docs.
A few things jump out. Hyperbrowser starts every session with “a fresh user data directory” and persistChanges defaults to false (Hyperbrowser), which is a safe default, but its docs describe no credential feature and say nothing about retention or encryption. Anchor’s plain profiles “do not store credentials or re-authenticate automatically”; you need its managed identity for that (Anchor).
Skyvern’s login call waits up to 1,800 seconds by default for the login to finish, and can take codes from a totp_url or totp_identifier (Skyvern SDK). Cloudflare stores no credentials at all, which I read as a feature: nothing to leak once the session ends.
If you’re weighing the two vendors I’d pick for logins, my Browserbase vs Steel comparison covers the rest of the trade-offs.

Is the site OK with an agent logging in? #
Some now say no in writing, and none of the five guides I read from the top of the search results mentions it. Before you connect an account, check the site’s terms; this section is a summary of public documents, not legal advice.
| Site or case | What it says | When |
|---|---|---|
| eBay User Agreement | No “buy-for-me agents, LLM-driven bots, or any end-to-end flow that attempts to place orders without human review” without “prior express permission” | Effective June 28, 2026 for existing users |
| United | It will “prohibit use by unauthorized, third-party personal AI agents or assistants”; sharing credentials with them “can expose sensitive information” | Announced November 11, 2026 |
| Bank of America | “If you do share your credentials, keep in mind that you’re responsible for activity resulting from that access.” | Customer email reported October 28, 2026 |
| May block sign-in from browsers “controlled through software automation rather than a human” | Policy since 2020, still on the help page | |
| Amazon v. Perplexity | Ninth Circuit vacated the preliminary injunction on anti-hacking-law grounds, as reported | August 4, 2026 |
Verified November 17, 2026.
United cites cybersecurity and federal aviation rules that require a human at check-in, according to TechCrunch, which also noted Bank of America’s warning the week before. I couldn’t load United’s own help and legal pages, so I don’t know the contract wording or how it will enforce it; my post on United’s agent restrictions follows that story, and eBay’s buy-for-me agent clause has its own post.
Google’s block goes back to a 2020 developer post that says “The browser must not provide automation features.” Google’s help page tells app developers to “Switch to use browser-based OAuth.” That matches what builders report: in an OpenAI forum thread about an agent that couldn’t sign in to Google, user Chris_De_Vargas wrote “it only works in the demo” (OpenAI community).
On the court case, AI Weekly reports the Ninth Circuit wrote that “it is Amazon’s customers—not Perplexity—that access Amazon’s computer systems” and called the Computer Fraud and Abuse Act “principally an anti-hacking statute”, while reserving the question of more autonomous agents. I didn’t read the opinion itself. My reading: a ruling about the anti-hacking law doesn’t settle what a site’s own terms allow.
Account risk is the first thing people ask about, too. On a November 2026 Launch HN for an agent product, user theturtletalks noted that “some say it gets your account banned” (Hacker News). I can’t tell you how often that happens; I can tell you which sites have written it down.
Also remember the site can see the agent. Cloudflare says Browser Run requests are “always identified as bot traffic” (Cloudflare), and signed agent identity is where this is heading; my Web Bot Auth explainer covers it. My rule: read the terms, prefer the site’s OAuth or API, and leave accounts that forbid agents alone.
Logged in makes prompt injection worse #
A signed-in agent that reads a hostile page can be steered to act with your privileges. Anthropic’s computer use docs advise against giving the model sensitive data such as account login information, and say a login “increases the risk of bad outcomes as a result of prompt injection”.
OWASP’s LLM01 entry gives the remedy as least privilege: “Provide the application with its own API tokens” and handle those functions “in code rather than providing them to the model”, with a human approving privileged operations. OpenAI’s guide adds a fresh browser profile or isolated desktop, and confirmation before consequential actions, including typing sensitive information.
So every login pattern above still needs confirmation gates on sends, purchases and settings changes. I compare the defenses in which prompt injection guardrails fit your deployment.
What about ChatGPT, Claude and Chrome? #
Each one documents a way to keep the password itself away from the model. If you’re handing a logged-in task to a consumer agent rather than building one, the vendor makes most of these choices for you. Here’s what each documents.
| Agent | How sign-in works | What the model sees | The catch |
|---|---|---|---|
| ChatGPT cloud browser | Pauses at sign-in; you type into a secure form; may ask for a 2FA code | Credentials are “not visible to the model”, and ChatGPT doesn’t store them | The session persists “until it expires”; paid plans except Free and Go |
| Claude in Chrome | Runs in your Chrome; 1Password for Claude can fill logins | With 1Password, passwords and codes “never enter Claude’s context” | You’re “responsible for all browser actions” |
| Chrome’s agent | Asks you to confirm before signing in via Google Password Manager | “the model does not have direct access to stored passwords” | You can pause and take over from a work log |
Verified November 17, 2026.
OpenAI’s help page also says a review model checks sign-in pages for phishing, and it’s direct about the rest: “Never paste passwords, security codes, or payment details into the conversation.” Sign-in arrived on August 25, 2026, after earlier docs had ruled it out, per Notebookcheck.
Anthropic’s advice for Claude in Chrome is a separate browser profile “without access to sensitive accounts (such as banking, healthcare, government)”, and Claude “asks for permission before accessing financial sites”. I’d follow both: a clean profile with only the accounts that task needs.
If an older guide tells you how ChatGPT Atlas handles logins, it’s out of date. OpenAI’s help page is now titled “ChatGPT Atlas retirement” and tells users to “Sign in to websites again in your new browser” (OpenAI); The Next Web reported an August 9, 2026 target for the shutdown.
Which login setup should you pick? #
Default to a login you make yourself in a saved per-site profile, then match every other account to one of these.
- Your own dashboards, portals and shop admin: log in yourself once through live view into a Browserbase Context with
persist: true, one per site and per user, then reuse it read-only. It’s encrypted at rest and stores no password. Browser Use or Steel profiles work the same way if you’re already on them. - The agent must re-login unattended: Steel’s Credentials API, because it’s origin-bound and blurs fields by default, generates codes server-side, and costs nothing during the beta (verified 2026-11-17). If beta status or org-wide storage is a deal-breaker, use Kernel’s or Anchor’s managed auth.
- A person should approve each login: 1Password for Agents on Browserbase, plus your own per-task approval rule, because one 1Password grant lasts 30 days and can cover multiple fills.
- The site offers OAuth, an API or an MCP server: use it over any browser login. You get scopes, expiry and revocation, and Google tells developers to do exactly this.
- Banking, airline loyalty, eBay buying: don’t give the agent the login. eBay and United restrict agents in writing, and Bank of America says you’re liable.
- No code: ChatGPT’s cloud browser sign-in form, or Claude in Chrome in a separate profile with 1Password for Claude.
The reason I default to a hand-made login over a stored password: a session expires and can be signed out on the site, while a stored password plus 2FA seed lets an agent, or whoever compromises it, keep coming back.
A checklist before the agent logs in anywhere #
- One profile per site and per user. Never share a Context or profile across customers or accounts (Browserbase, Browser Use, Cloudflare).
- Add a logged-out check. Detect the sign-in page and stop, rather than letting the agent improvise (Browserbase).
- No credentials in prompts. Use type-time secrets or a password manager (Steel, OpenAI).
- Assume screenshots see everything. Turn on masking where it exists; Browser Use says screenshots aren’t redacted (Browser Use, Skyvern).
- Keep live-view links short-lived and private. Treat them like passwords (Cloudflare, Steel).
- Gate every state change. Publishing, payments, settings, deletions and permission grants wait for a person (BrowserAct, OWASP).
- Know how to revoke. Signing out on the site, rotating the password and 2FA seed, deleting the profile; deleting the profile alone isn’t enough (Browser Use).
- Read the site’s terms. If they forbid agents, leave the account out (eBay, United).
FAQ #
Is it safe to give an AI agent my password? #
Not in the prompt. OpenAI tells ChatGPT users never to paste passwords into the conversation, and Anthropic’s docs say a login increases the risk of prompt injection. Log in yourself into a saved profile, use a vault that types the secret at fill time, or use OAuth where the site offers it. For banking, Bank of America told customers that if they share credentials with an agent, they’re responsible for what follows.
Can a browser agent handle two-factor codes? #
Yes, if you store the authenticator seed with it: Steel, Browser Use, Kernel, Anchor and Skyvern all document that. But then the agent holds both your password and your second factor, so 2FA no longer shows a person was there. For sensitive accounts, hand the 2FA step to a person through a live view instead.
Is a saved browser profile as sensitive as a password? #
In practice, yes. Playwright warns that saved browser state could be used to impersonate you, and MITRE ATT&CK lists stolen session cookies as a way in that needs no credentials and can get around some multi-factor checks. Store profiles like secrets, one per site and per user, and know how long your vendor keeps them.
Can a browser agent use my passkeys? #
Not on its own. FIDO says passkey user verification can only be done by the user with a biometric or PIN. Skyvern lists passkeys as an Enterprise feature that cannot use saved browser sessions, and FIDO’s agentic authentication working group, formed in April 2026, named no spec dates when it launched.
Could a site ban my account for using an AI agent? #
I can’t tell you how any site enforces its rules, but some now restrict agents in writing. eBay’s User Agreement bars buy-for-me agents and LLM-driven bots without permission, United says it prohibits unauthorized third-party AI agents, and Bank of America says you’re responsible if you share your login. Read the terms first; this is not legal advice.
Why won’t Google let my agent’s browser sign in? #
Google says it may block sign-in from browsers controlled through software automation rather than a human, and tells app developers to switch to browser-based OAuth. Google sessions on Chrome for Windows are also now bound to the device, so a copied session can stop working elsewhere. Use OAuth or an API for Google accounts.
Does one 1Password approval cover every login the agent makes? #
Not every login, but often more than one. 1Password’s partner docs say a grant lasts 30 days at launch and one approval can cover multiple fills; approving each session is up to the platform that integrates it. If you want a person to approve every task, build that rule yourself.
Sources
55 sources, accessed .
Official pricing, docs, changelogs and status pages 41
- Browserbase docs, Contexts
- Browserbase docs, 1Password quickstart
- Browserbase changelog, 1Password for Agents (Oct 22, 2026)
- Browserbase, prompt injection containment (Feb 13, 2026)
- Browser Use Cloud docs, authentication
- Browser Use docs, profiles
- Browser Use docs, profile sync
- Browser Use open source docs, authentication
- Browser Use open source docs, sensitive data
- Browser Use Cloud docs, 1Password
- Steel docs, Credentials API overview
- Hyperbrowser docs, profiles
- Anchor docs, secret values
- Anchor docs, managed authentication
- Kernel docs, Managed Auth overview
- Kernel docs, Managed Auth (TOTP)
- Skyvern SDK reference, agent login
- Skyvern docs, credentials overview
- Cloudflare Browser Run docs, human in the loop
- Cloudflare Agents docs, browser tool
- 1Password, Closing the credential risk gap for browser-use AI agents (Oct 8, 2025)
- 1Password developer docs, agentic autofill for partners
- 1Password developer docs, 1Password for Agents
- Playwright docs, authentication
- MITRE ATT&CK T1539, Steal Web Session Cookie
- Google, Device Bound Session Credentials (Apr 9, 2026)
- Google Workspace Updates, DBSC generally available on Chrome for Windows (May 2026)
- Google Account Help, browser may not be secure
- Google Developers Blog, blocking less secure browsers (Aug 28, 2020)
- Google, Architecting security for agentic capabilities in Chrome (Dec 8, 2025)
- Claude computer use tool docs
- Claude Help Center, use Claude in Chrome safely
- OpenAI computer use guide
- OpenAI Help, using cloud browser in ChatGPT (read via reader proxy)
- OpenAI Help, ChatGPT Atlas retirement (read via reader proxy)
- Model Context Protocol, authorization (2025-11-25)
- OWASP LLM01:2025 Prompt Injection
- FIDO Alliance, passkeys
- FIDO Alliance, standards for trusted AI agent interactions (Apr 28, 2026)
- eBay User Agreement
- United, United Restricts Access for Third-Party AI Agents (Nov 11, 2026)
Community discussions 3
News and press 5
- TechCrunch, United Airlines blocks AI agents (Nov 12, 2026)
- Android Authority, Bank of America AI agent password warning
- AI Weekly, Ninth Circuit lifts CFAA injunction on Perplexity's Comet (Sep 10, 2026)
- Notebookcheck, ChatGPT can now sign in to your accounts (Aug 26, 2026)
- The Next Web, OpenAI's Atlas shutdown (Jul 10, 2026)
Other blogs 6
- Steel, secure browser auth for agents (Mar 31, 2026)
- Browser Use, web agent authentication (updated Oct 25, 2026)
- WorkOS, logging AI agents into web apps (Mar 31, 2026)
- Logto, Manus cloud browser login (Jun 27, 2025)
- BrowserAct, AI agents and logins (Sep 10, 2026)
- Browserless, the problem with persisting profiles (Aug 18, 2026)


