OpenAI’s October 15 Chrome extension update expands what ChatGPT can do with open browser tabs, according to Engadget’s October 16 report. Chat can use information across tabs to answer questions; Work can carry out browser tasks such as filling forms, sending messages and organizing tabs. The update extends those capabilities rather than establishing that Chat and Work are newly created modes.
Unite.AI’s account of the announcement also describes tab grouping and a sidebar with conversation history. Its summary of the store listing says the extension asks for permission when using browser history and when uploading or downloading files. We treat those as reported controls, not as independently tested behavior or a promise that every action always triggers a confirmation.
The retrieved OpenAI browser-extension guide confirms that the desktop app can work with already signed-in sites. It lists tab mentions and browser control for Chrome, Edge, Brave, Opera and Vivaldi, with side chat unavailable in Opera. It also warns that availability can depend on rollout and workspace settings. That is a narrower statement than assuming identical functionality for every account and browser.
The guide describes using open-tab context in side chat, continuing those conversations in the app, and treating page content as untrusted. Those details matter to the practical distinction: having a page available as context and authorizing the assistant to act on that page are separate decisions. We have not performed a hands-on test of the announced update and do not infer access to all tabs, all sites or every workspace feature.
Our analysis: a useful first evaluation would separate a reading task from an action task. For a hypothetical supplier comparison, the reading task could ask for delivery terms from three open pages, with links back to the evidence. A reviewer could check whether each term actually appears on the cited page. That would test the answer produced from context; it would not establish that the assistant can place an order correctly.
The action task would need its own boundary. An evaluator could ask for a draft form to be completed while requiring the process to stop before submission. The review should inspect the fields, the selected site and whether the requested stopping point was respected. This is a proposed test, not a result from using ChatGPT, and success on one form would not prove reliability across unrelated websites.
Existing sign-in makes that separation especially relevant. The primary guide explicitly describes work on authenticated sites. In a proposed workplace pilot, we would therefore verify which account is open before granting a task and keep the requested outcome specific. We would not assume that the presence of a logged-in session grants permission to make every change that account can technically perform.
There is also a difference between remembering a conversation and having authority over a website. The reporting describes conversation history in the sidebar, while the guide explains how side-chat conversations can continue in the app. Neither point establishes unrestricted permission to act on an external system. A familiar conversation should still be evaluated against the particular task and website involved.
For a team assessing the update, we would record three things separately: the pages supplied as context, the actions requested and the observable result. If an answer is wrong, that record helps distinguish an incorrect reading from a wrong site selection or an action that exceeded the request. It would also keep a successful text summary from being mistaken for evidence that a multistep workflow completed correctly.
The practical value of this release is the reported connection between browser context and tasks performed alongside a website. Its usefulness will depend on the actual workflow, account settings and browser combination. The sources establish the announced functions and documented scope; they do not establish a comparative productivity gain, universal rollout or a blanket privacy guarantee.
Before delegating a browser workflow, verify your browser and workspace support, specify which tabs to use, and set an explicit stopping point before any submission you intend to review.
