We spoke to browser agent developers and saw people copying and pasting credentials and even credit card numbers directly into model system prompts. We were surprised that there wasn’t a better way to give agents access to websites on a human’s behalf. Moreover, we noticed that every company had to build infrastructure to manage OTP, TOTP, and MFA and that auth remained a significant hurdle in agent reliability. We wondered if this was a boring part of the problem of building web automations that someone could automate away.
We started working with Casco, an autonomous security testing company, to enable their agent to access customer sites. Before a pentest, Casco makes a request to Prism’s API specifying test user credentials, a domain, and a login method. For example, give me an authenticated session for the account rajit@prismai.sh for OpenAI via OTP code over email. Our agent logs in on their behalf (without exposing credentials to a model), and we download the cookies and send them back in the response.
To maintain speed and reliability, we use playwright in most cases to login (which gives us speed), and we fallback to AI on failure (which gives us reliability). We have a number of websites we support out of the box and add new scripts as the number of websites we need to support grows. We are working on a way for the agent to update the existing playwright script on failure, so our scripts always stay up to date.
To try our api, you can use our API playground docs.prismai.sh/api-reference/endpoint/login to sign into x.com with the following API key: pk_54abb1cd0a637eb973ed690416e71a953e98f2ea839cf16529bbfa41a41bc016 .
We’d love to learn more about how other developers give agents access to their accounts. We look forward to everyone’s feedback and comments.
