Cuiar
Sign inDownload

Security

Security, in plain words.

The main points first. The details follow. What we keep and for how long: Privacy.

In short

  • The agent asks you before it pays, posts, sends a message or changes an account.
  • Text from websites is data. It cannot give the agent orders.
  • The AI model never sees your passwords. The vault types them into the page.
  • The browser has no Google sign-in, no Safe Browsing calls and no usage statistics.
  • You sign in with a one-time code by e-mail. We keep only hashes of tokens and keys.
  • Report a vulnerability to security@cuiar.com.

The agent

It asks before it acts

  • Cuiar's agent asks you before it pays or buys, posts, sends a message or changes an account, for example a new password or a deleted account.
  • The browser finds these actions itself, so the agent stops for your answer also when the model forgets to ask.
  • You answer in the side panel. For a task that you started on cuiar.com, in the CLI or from a routine, you can also answer by e-mail or in the CLI.
  • When no one answers during a routine in 10 minutes, the agent does not do the action.

Page text is data, not orders

  • Text from a website goes to the model between markers with a random ID. The page cannot write or close these markers.
  • The model is told never to follow requests, commands or rules in that text, and never to let it change the task.
  • Old page text leaves the request after two newer page reads, so a page cannot stay in the agent's context for the whole task.

Passwords stay out of the model

  • The vault types the user name, the password and the one-time code into the page. The model gets only "signed in as" and the user name.
  • The vault fills a login only on that login's site. It checks the site's address first.
  • Set a login, or the whole vault, to Never, and the agent cannot use it. Incognito tasks never use the vault.
  • Each use of a login goes into the vault's activity log on cuiar.com/vault.

Limits that you set

  • Read only: the agent reads pages and changes nothing.
  • Guard: you approve each change that the agent makes. This is the default for tasks that you start on cuiar.com, in the CLI or with the API.
  • A task stops after 40 steps, or 120 for Odyssey, and waits for your next message.
  • Pause or stop a task at any time.

When your own AI drives Cuiar

  • Claude, Codex or another MCP client can use Cuiar's browser tools. That client's own approval settings apply to its calls. Cuiar does not ask again.
  • Page text from the tools is marked as untrusted data in the same way.
  • The vault still signs in without showing the password to the client.

The browser

Chromium 155

Cuiar is built on Chromium 155.0.8059.79. In 2026 its base moved to each new major Chromium version, from 144 in February to 155 in October.

  1. 144Feb
  2. 145Mar
  3. 146Apr
  4. 147May
  5. 148May
  6. 149Jun
  7. 150Jul
  8. 151Aug
  9. 152Sep
  10. 153Sep
  11. 154Sep
  12. 155Oct

Google services and telemetry, removed

No Google account in the browser
Cuiar has no Google API keys, so the browser cannot sign in to a Google account or sync with Google. Google websites work as usual.
No Safe Browsing
Cuiar does not send the addresses that you open to Google. It also does not show Google's warnings about unsafe sites.
No field trials
Cuiar does not download Google's experiment settings, and it does not add the X-Client-Data header to requests.
No usage statistics or crash uploads
Their Google addresses are removed. The crash handler is off on macOS and Windows.
No other Google calls
Google update checks for components and extensions, network time, Translate, promotion tracking (RLZ), surveys and Domain Reliability reports are off or have no address.
No Google addresses
The build changes the Google host names in the Chromium source to names that do not exist. A Google request that stays in the code goes nowhere.

Our servers

The site, the API, the MCP server and the bridge that connects Cuiar on your computer to them.

  • cuiar.com is a static site on Cloudflare. Its Content Security Policy lets only scripts from cuiar.com run, so the site loads no third-party script.
  • api.cuiar.com is behind Cloudflare. The API server answers only requests that come through Cloudflare.
  • You sign in with a one-time code by e-mail, so there is no account password to steal. A code expires in 10 minutes, and sign-in attempts have limits per IP address and per e-mail address.
  • A site sign-in lasts 30 days, and an app or CLI sign-in 90 days. Sign out any device on cuiar.com/account.
  • We store each session token, API key, license token and OAuth code only as a SHA-256 hash.
  • AI clients connect with OAuth 2.1 and PKCE (S256), after you approve each one on cuiar.com. API keys have scopes.
  • Task files are in a private bucket. A download link works for 10 minutes, and a link in a data export for 24 hours.
  • Routine results that go to a webhook carry an HMAC-SHA256 signature.
  • Cuiar on your computer connects out to our bridge. The bridge never connects in to your computer.
  • Stripe takes payments. We never see a card number.

Vulnerabilities

Report a vulnerability.

security@cuiar.com

Send what you found, the steps to see it again, and what an attacker can do with it.

Test only with your own account and your own data. Do not read, change or delete the data of other people, and do not send traffic that slows the services down.

Our contact data for security is also at /.well-known/security.txt.