Beginner Basics

The model should never have access to your password.

Hyperbrowser credential fill encrypts secrets, keeps passwords and API keys out of LLM context, and types them only into allowed browser origins.

The model should never have access to your password.

Key Takeaways

  • Hyperbrowser credential fill keeps passwords, API keys, TOTP seeds, and 1Password tokens out of the LLM context by storing them encrypted and typing them directly inside the browser session.
  • It reduces leakage paths by restricting fills to allowed origins, separating filling from submitting, and returning detailed failure outcomes without typing values.
  • Models can reference secret names and CSS selectors, but never see the actual values; navigation, clicking, and form submission remain under developer control.
  • It supports 2FA and rotation through TOTP secrets, secret versions, and 1Password vault references, while avoiding persistent storage of 1Password values.
  • Residual risk remains: once a secret is typed into a page, site scripts, extensions, CDP access, recordings, and visible fields can potentially expose it.

Meta description: Hyperbrowser credential fill stores secrets outside model context, types them in browser sessions, and uses allowed origins to reduce password leakage.

The model should never have access to your password.

Tags: agent security, credential management, browser automation, Hyperbrowser, LLMs

!Abstract illustration of encrypted credentials entering a browser session without passing through a model context.

This review, based on the October 2026 launch post and documentation, is for developers and agent builders who use browser automation and want to understand how Hyperbrowser credential fill keeps passwords, API keys, TOTP seeds, and 1Password service account tokens out of LLM context. It explains the leak paths it removes, the session flow, the limits that matter, and where the feature still leaves risk.

Why the Demo Password Matters

Browser agent demos often show the same sequence: an agent types an email, types a password, clicks through a two-factor screen, and reaches a dashboard. It looks automatic, but browser agents expose the worst part of the workflow. To type a password into a login form, something in the stack must know the password. Today, that something is usually the LLM.

That value can appear in a prompt, a tool call argument, the transcript, application logs, traces, and the request history kept by a model provider. None of these are exotic bugs. They are normal side effects of asking a text model to be the courier for a secret.

The model should never have access to your password.

How Hyperbrowser Credential Fill Changes the Path

Hyperbrowser credential fill removes the model from that path. You create a secret through the SDK, the hx CLI, or the dashboard. The SDK encrypts the value on your machine before upload, and the API stores the encrypted envelope. A secret can be a password, API key, base32 TOTP seed, or 1Password service account token. You also attach allowed origins, the only sites where it may be typed.

When you create a browser session, you attach the vault or individual secrets by name. A session can fill only what it was created with, up to 20 credentials. Attaching secrets forces a Firecracker browser and turns off Chrome's password manager. Once the session is on a login page, one fill call can fill the whole form; each field names a vault entry and a CSS selector. The platform decrypts each secret inside the browser VM and types it in. Values never appear in code, logs, or the model context.

The response reports, per input, an outcome and the secret version used. filled means the input holds the stored value, not_found means no input matched, ambiguous means multiple visible inputs matched, not_editable means the input is disabled, read-only, hidden, or not a text field, and not_accepted means the page did not keep the value, such as because of a maxlength. Only inputs a person could see count, so hidden copies such as mobile layouts or honeypot fields are skipped.

A fill can take up to about 18 seconds, so an HTTP client should allow at least 30 seconds. It works with Playwright or Puppeteer scripts that connect to the session over CDP; navigation, clicking, and submitting stay in your code.

Compared with prompts, tool arguments, or environment variables—where the value can stay in transcripts, request history, logs, or process lists, and rotation means re-prompting, redeploying, or rebuilding images—credential fill keeps the model away from the value, keeps it out of logs and traces, makes rotation an API call, and restricts typing to allowed origins.

The model should never have access to your password.

Design Choices That Reduce Risk

  • Credential fill types values but never submits. A fill that submits could happen at the wrong moment or on the wrong page, with no way back. A fill that only types is safe to retry and leaves the submit decision to your code or the model.
  • Allowed origins live on the secret, not the request. If the target origin is wrong, the fill fails with CREDENTIAL_ORIGIN_NOT_ALLOWED and nothing is typed. The rules are strict: 1 to 20 origins in scheme://host[:port] form, no paths, wildcard subdomains that match subdomains but not the apex domain, and http:// accepted only for localhost. This means a compromised request cannot widen where the value may go.
  • The fill_credential tool definition for OpenAI, Anthropic, and Gemini lets the model name a secret or vault and pick selectors, but it never sees values. It deliberately does not support 1Password fills, because a 1Password fill names allowed origins in the request, so letting a model choose them would let the model choose where values may be typed.
  • The error taxonomy assumes failure is normal. It covers multiple tabs with CREDENTIAL_PAGE_AMBIGUOUS and candidate details, iframe login forms via frameUrl where the iframe's own origin must be allowed, pages restored from a snapshot with Chrome's password manager on, and malformed TOTP seeds. Nothing is typed in any failure.
  • The docs admit what stays readable. Once a value is in the page, site scripts, browser extensions, and your own CDP connection can read the field. Video recordings and the live view show what is on screen. Password inputs are masked by the page, but a filled username or one-time code is visible. The feature takes the secret out of code, logs, and the model, but it does not claim to protect the value after it lands in a page.

One-Time Codes, 1Password, and Rotation

For two-factor logins, you store the authenticator's base32 seed as a secret of type totp. Each fill types the current code, and the result includes an expiresAt value so you know when to fill again before submitting. The seed does not leave the browser VM.

For teams whose logins already live in 1Password, a session can fill fields read directly from a 1Password vault using a service account token. The values are read at fill time and are not stored by Hyperbrowser. Each field uses a reference such as `op://Dev Vault/scrapingcourse login/username`. One-time codes use an `?attribute=otp` suffix.

Rotation is handled through secret versions. You store a new version, and fills use the latest version unless you pin one. Credential fill is in public preview and enabled for teams on request. Teams need to contact Hyperbrowser support to enable it. A restricted API key needs secrets.manage to create secrets and secrets.use to fill them.

Who should use it / who should skip it

Use Hyperbrowser credential fill if you build browser agents that sign into real accounts and need to keep passwords, API keys, TOTP seeds, and service account tokens out of prompts, traces, logs, and model context. It is useful when you can define strict allowed origins, accept a fill-only step before submission, and treat page-level exposure as a remaining risk.

Skip it, or delay it, if you expect the model to choose 1Password origins, need Chrome's normal password manager behavior, cannot control selectors and page states, or are not ready to review what site scripts, extensions, CDP access, video recordings, and live views may expose after a field is filled. The feature removes the model from the credential path, but it does not make a logged-in page safe by default. The design's value is restraint: it declines model capabilities in at least three places instead of making the demo look more flexible.

Comments (0)

  1. No comments yet. Be the first to share what worked for you.

Leave a comment

Comments are reviewed before they appear. Your email address is not published.