Between October 15 and 18, 2026, OpenAI deployed a major update to its ChatGPT extension for Google Chrome (Engadget). The software, which listed approximately 7 million users and a 2.2-out-of-5 rating across 922 reviews on the Chrome Web Store as of October 16, 2026 (Tech My Money), now incorporates two distinct operational modes: Chat mode and Work mode (Yahoo Tech).

This development transitions capabilities previously confined to OpenAI's discontinued Atlas browser—which shut down on August 9, 2026 (International Business Times Australia)—directly into the most widely utilized web browser ecosystem (Innovation Village).

Dual-Mode Architecture: From Passive Analysis to Active Execution

The core structural shift in the extension lies in its division between passive information processing and active browser manipulation. In Chat mode, the assistant is configured to scan all open browser tabs to gather immediate context, summarize long-form articles, extract data points, and facilitate side-by-side product comparisons (Engadget). The redesigned side-by-side sidebar also now displays the user's complete historical conversation log alongside support for external plugins and local file system access (Tech My Money).

Conversely, Work mode grants the assistant active control over the browser interface (Yahoo Tech). Under this configuration, ChatGPT can execute multi-step routines such as filling out online forms, sending messages, downloading files, and programmatically organizing open tabs into categorized folders (International Business Times Australia).

These capabilities are accessible across all subscription tiers, including the free plan (Innovation Village), and function natively on Chrome as well as alternative Chromium-based environments like Microsoft Edge, Brave, Opera, and Vivaldi (Engadget).

Permission Models and Privacy Safeguards

Because the extension reads across active windows and executes commands, OpenAI has implemented a permission structure intended to mitigate security risks. The system relies on step-by-step explicit user confirmation before the AI can access new websites or execute critical interactive tasks (Tech My Money). Furthermore, users retain the ability to configure granular domain-level permissions using allowlists and blocklists (Innovation Village).

Despite these controls, the architectural reality of an AI extension sweeping across active browser sessions introduces surface area for unintended data exposure. OpenAI advises users to manually close sensitive tabs—such as those containing banking information or personal health records—before delegating automated tasks to the assistant (Innovation Village). As browser-based agents transition from theoretical demonstrations to deployed desktop capabilities, the operational friction between continuous ambient context gathering and user privacy remains an active engineering and governance challenge.