OpenAI has updated the ChatGPT browser extension for Chrome to version 1.26.1005.11759, turning the assistant into a full-fledged browser agent. With the @chrome command, it reads the content of open tabs and extracts data from multiple pages in parallel, while the new Work mode, based on Computer Use technology, performs actions: filling out forms, downloading files, sending messages, and grouping tabs. The update, dated October 15, 2026, is available on all ChatGPT plans and works in Chrome, Edge, Brave, Opera, and Vivaldi, with the extension already having approximately 7 million installs in the Chrome Web Store. The mass browser agent has become a reality, but there are no independent measurements of its actual reliability yet.

image

What happened

OpenAI released an update to the ChatGPT browser extension for Chrome, version 1.26.1005.11759, dated October 15, 2026. The @chrome command gives the assistant access to the content of open tabs: the model can extract data from multiple pages in parallel, for example, comparing prices from different sellers or summarizing long articles. A Work mode has been introduced for practical tasks, built on Computer Use technology: the assistant fills out forms, downloads files, sends messages, and groups tabs. The updated extension is available on all ChatGPT plans and works in five Chromium-based browsers — Chrome, Edge, Brave, Opera, and Vivaldi. The full conversation history is displayed in the updated sidebar.

Context

To assess the scale of the shift, it is worth recalling the history of Computer Use: the technology for controlling interfaces at a human level existed at OpenAI as isolated demos until recently. The company also developed a separate browser, Atlas, as an independent product, but it was discontinued, so the path of an "agent on top of an already installed browser" has now been chosen. This task is solved by the extension's architecture: it requests broad permissions, including access to data from all sites, history, bookmarks, and the debugger API, and on top of this, specific runtime prompts are formed for each site. By nature, this is news about a deployment, not new science: the update combines known components in a new package, rather than introducing a new architecture, training method, or research breakthrough.

Why this matters for the industry

For the AI industry, the main point here is not the set of features, but distribution and competitive dynamics. With approximately 7 million installs in the Chrome Web Store and availability on all ChatGPT plans, Computer Use has reached a mass audience for the first time, rather than remaining in demonstrations, so the browser agent is transforming from a niche feature into a standard interface for AI automation of web tasks. Simple web automation is being commoditized at the consumer level today: the value of lightweight products around one-time data extraction from tabs is being washed out and shifting toward the orchestration of more complex agent scenarios. For teams building products, this is both an opportunity — prototypes of agent flows can be assembled in hours without their own backend, and the browser becomes a ready-made channel for distributing agent functions. Another direct consequence concerns web development: an extension with broad permissions and individual runtime prompts for each site creates a class of prompt injection risks, and web application developers have to consider AI agents as legitimate users of their interfaces.

Why this matters for users

Users can try it immediately: the update has already been rolled out to all ChatGPT plans, and no setup is required for the first steps. A reasonable order is from reading to actions: read-only tasks are performed in Chat mode, while Work mode, where the agent acts independently, should be tried only on non-critical sites where an erroneous action will not break anything. Separately on permissions: for each new site, the extension requests consent separately, offering options such as Allow once, Allow for this site, and Allow for all sites, so control remains with the user. It is not advisable to grant global access to all sites immediately, and it is better to remove API keys and passwords from open tabs while the agent is working: the assistant reads the content of tabs in full.

What is still unknown / limitations

The main caveat is the lack of measurements. The sources contain no measurements of success rate, data extraction accuracy, or latency; no one has checked whether forms are filled out stably or how often the agent makes mistakes, so claims about Work mode capabilities remain vendor claims rather than measured facts. The evidence base consists of the OpenAI announcement and secondary coverage — articles from The Decoder and the explainx.ai blog — without methodology, metrics, or reproducibility. The extension has no public API, per-action pricing, or limits, and without these mechanisms, it is not yet possible to give the agent production access, payment, or message sending on behalf of a company. The risk of prompt injection is mentioned in the sources, but data on the agent's resistance to it has not been published. Expectations that prompt injection will become an active area of attacks and research in six months, and that the browser agent may become a standard interface for web automation in two years, are interpretations, not established facts.

Sources

Author

Look at AI, editorial team