Who can start a run
- Only a signed-in member of your organization can start a run.
- Before a run starts, we check its target address. Private, loopback and cloud metadata addresses are refused.
Your logins
- The agent only logs in to your site if you give it a login, either for the project or for a single run.
- A saved project login is encrypted with AES-256-GCM before it is stored.
- The planning model is never given the password. It leaves the password field empty, and our runtime types the password in.
- The screenshot after the step that types a password is not saved.
- A saved project login is never used on a pull request from a fork.
What the agent can do in a browser
- It can click, type, hover, drag, scroll, press a key, pick from a dropdown, go to a URL, wait, and check the screen. It has no action to upload or download a file.
- Around steps that submit, pay, order, send or delete, it checks the result more strictly, and it will not repeat a step that may already have worked. A payment is not made twice.
- Each run has hard limits: 270 seconds by default, 45 seconds per step, at most 100 steps, at most 3 re-plans, and a model cost ceiling of $1 per run in production.
What you can check after a run
Every step records a screenshot before and after (except the after shot of a password step), the reasoning of the model that acted, and the reasoning of the model that checked the result. All of it is on the run page, so you can see why a step passed or failed.
Your code and pages sent to models
- Every model call is restricted, in code, to zero-retention endpoints that do not train on what we send.
- We do not capture model prompts in production.
Page checks and the free scan
- The accessibility, performance, SEO tag, console and network, and cookie checks load your page once and read it.
- The free scan on our home page, which needs no account, runs only read-only page checks. It does not click, type, fill a form or log in.
- The free scan refuses login, account and checkout paths.
- The free scan is limited to 3 scans an hour and 10 a day per IP address, and 100 a day in total.
Report a vulnerability
If you find a security issue, please tell us privately at [email protected] before you share it publicly. We will reply, look into it, and keep you updated on what we find.