# Can the Agents API hosted browser fill sensitive fields without exposing them to the model?

**URL:** <https://community.openai.com/t/can-the-agents-api-hosted-browser-fill-sensitive-fields-without-exposing-them-to-the-model/1417137>\
**Category:** API\
**Tags:** agents-api\
**Created:** [November 16, 2026, 8:18pm UTC](https://community.openai.com/t/can-the-agents-api-hosted-browser-fill-sensitive-fields-without-exposing-them-to-the-model/1417137 "2026-11-16T20:18:37Z")\
**Posts on this page:** 1\
**Page:** 1

<div class="post-metadata">

**Author:** ![plazari](https://sea2.discourse-cdn.com/openai1/user_avatar/community.openai.com/plazari/32/776506_2.png) [@plazari](https://community.openai.com/u/plazari)\
**Post date:** [November 16, 2026, 8:18pm UTC](https://community.openai.com/t/can-the-agents-api-hosted-browser-fill-sensitive-fields-without-exposing-them-to-the-model/1417137/1 "2026-11-16T20:18:37Z")

</div>

Hi everyone,

I’m building a PoC with the Agents API, the OpenAI-hosted browser (`computer_use` with `environment.desktop.enabled`), and some custom MCP tools.

The agent navigates a shopping website and helps the user through checkout. When it reaches a form asking for sensitive information, I’d like to collect that information in my application’s UI and fill it in the same browser, without sending the values to the model.

For login, I found the `browser_authentication` flow. According to the docs, values submitted through its dedicated event stay outside model input and are omitted from authentication response items in session history.

Is there something similar for fields that aren’t part of login? For example, a Brazilian tax ID (CPF) requested at checkout, or payment-card details?

Ideally, the agent would identify the fields, my UI would collect the values and the user’s permission, and the browser would fill them. The agent would only receive a success or failure status.

I haven’t found a documented way to do this. Returning the values through an MCP tool would expose them to the model, so that wouldn’t solve it.

Does the hosted browser support this today? Could a custom MCP tool fill fields in that same browser through a supported interface, or would I need to manage my own browser runtime?

I’m also wondering what happens after filling: could the values become visible to the model through subsequent screenshots or other browser observations?

I understand that the website and browser infrastructure need to process the data. I’m specifically trying to keep it out of the model context and conversation history.

Has anyone built something like this, or found a recommended approach?

Thanks!
