TechRadar reported on October 22, 2026, that ChatGPT Work paused during one webinar-registration form task to ask which industry category to choose. Eric Hal Schwartz had instructed it not to submit the form. The account describes one task, while OpenAI’s instructions explain how page-aware chat, browser actions and access permissions work.
Page context or browser control?
Chat uses page or tab information to answer a prompt; a Work or Codex task can interact with websites through a connected browser. OpenAI documents Chrome support for side chat, tab mentions and browser control from the ChatGPT desktop app.
| Workflow | What it does | How to start it | Website access |
| Chat and page context | Uses an open tab or selected page text as context for a response. | Open side chat or mention a tab in Chat. | Page content provides context; this is distinct from authorizing a browser task to interact with a site. |
| Work or Codex browser task | Can interact with a website, including a site where you are signed in, subject to access controls. | Start a Work or Codex chat and select the browser with an @ mention. | OpenAI says it asks before interacting with each new site by default. |
The distinction matters when you want a summary of several pages rather than an action on a site. In a creator demonstration, Igor Pogany shows setup and a query drawing on multiple open tabs.
For other browsers, OpenAI also documents tab mentions and browser control in Microsoft Edge, Brave, Opera and Vivaldi. Side chat is documented in Edge, Brave and Vivaldi, but not Opera.
Set up the desktop-app connection
OpenAI’s setup runs through the ChatGPT desktop app, rather than the browser extension alone. Follow the documented sequence:
- Install or update the ChatGPT desktop app.
- Open Settings > Computer Use.
- Select the browser and install the extension through that setup flow.
- Review the browser permission prompts.
- Confirm the browser shows Manage in Computer Use.
- In a Work or Codex chat, select the browser with an
@mention. Use the browser profile where the extension is installed.
OpenAI’s browser extension instructions also describe the browser-access controls and connection troubleshooting.
Review permissions before starting a task
Installation permissions and permission to use a particular website are separate choices. The extension permission prompt may include access to page data, browsing history, bookmarks, downloads, tab groups and the page debugger. OpenAI says ChatGPT asks before interacting with each new website by default; you can allow access once, allow it for that site, allow it for all sites, or decline. Computer Use settings also manage website allowlists and blocklists.
Allowing access for all sites removes the site-by-site confirmation, and OpenAI labels that choice elevated risk. OpenAI also labels browser-history access elevated risk. ChatGPT requests history access when needed, scopes it to the task, and does not offer an “always allow” option.
OpenAI advises treating webpage text, selected text and video transcripts as untrusted context and reviewing them before ChatGPT acts on them. That matters when a page contains instructions that you did not intend to give the assistant.
What the form test and CAPTCHA demonstration show
In the one form task described by TechRadar, Work paused to ask Eric Hal Schwartz which industry category to select for a webinar registration. He had told it not to submit the form. The account offers a specific example of a clarification pause, not a broad measure of performance across sites or tasks.
In a separate creator demonstration, Chris Winters shows a browser task that pauses when a CAPTCHA requires human intervention. It is one demonstrated task, not a guarantee of how every CAPTCHA or browser workflow will behave.
What OpenAI says about browser activity
OpenAI says it does not store a separate complete record of the extension’s browser actions. It says activity is stored when it becomes part of ChatGPT context, such as page text, screenshots, tool calls, summaries or messages. Browser access therefore does not mean that no task information enters ChatGPT context.
For background on the two Chrome workflows, NeoTeo previously explained what Chat and Work do in Chrome.
