ChatGPT Atlas Release Notes

OpenAI has introduced ChatGPT Atlas, a web browser that integrates ChatGPT directly into the browsing experience. This integration allows the AI to understand the user's active page context, remember details across different websites, and perform autonomous tasks via a native agent mode, eliminating the need for manual copying and pasting between the browser and a separate chat interface.

Integrated Browsing and Contextual Awareness

ChatGPT Atlas enables the AI to accompany the user across the web, providing assistance within the active window. Because the browser is built with ChatGPT at its core, the system can understand what a user is viewing in real-time to provide immediate help and knowledge checks without requiring screenshots or manual context sharing.

Browser Memories

Atlas introduces "browser memories," an optional feature that allows ChatGPT to remember context from visited sites to inform future interactions.

  • Functionality: Users can ask the AI to synthesize information from past browsing sessions, such as summarizing industry trends from job postings viewed in previous weeks.
  • Control and Privacy: Browser memories are private to the account. Users can view, archive, or delete these memories in settings. Deleting the browsing history also deletes all associated browser memories.
  • Page-Level Control: A toggle in the address bar allows users to disable ChatGPT's visibility for specific websites, preventing the AI from seeing page content or creating memories from that site.

Agent Mode and Task Automation

Agent mode allows ChatGPT to take autonomous actions within the browser to complete complex workflows. This feature is available in preview for Plus, Pro, and Business users.

Capabilities

ChatGPT can now perform tasks such as:

  • Shopping: Finding a grocery store, adding ingredients from a recipe to a cart, and placing an order.
  • Research: Reading through past team documents and performing competitive research to compile a team brief.
  • Execution: Opening tabs and clicking elements within the browser to fulfill a request.

Safety and Security Constraints

To mitigate risks associated with agentic access to logged-in sites, OpenAI has implemented several safeguards:

  • Technical Restrictions: The agent cannot run code in the browser, download files, install extensions, or access the local computer's file system or other apps.
  • Sensitive Site Protection: The agent will pause for user confirmation when taking actions on sensitive sites, such as financial institutions.
  • Logged-Out Mode: Users can run the agent in a logged-out mode to limit access to sensitive data.
  • Prompt Injection Risks: OpenAI acknowledges that agents remain susceptible to "hidden malicious instructions" embedded in webpages or emails that could attempt to override intended behavior to steal data or perform unauthorized actions.

Availability and Deployment

ChatGPT Atlas is available worldwide on macOS for Free, Plus, Pro, and Go users. It is available in beta for Business users and, if enabled by administrators, for Enterprise and Edu users. Versions for Windows, iOS, and Android are scheduled for future release.

Onboarding

New users can import bookmarks, saved passwords, and browsing history from their existing browsers during the initial setup process.

Data Privacy and Training

By default, OpenAI does not use browsed content to train its models. Users may opt-in to this via the "include web browsing" setting in data controls. However, the system respects the GPTBot opt-out signal from website owners; pages that opt out of GPTBot will not be used for training even if the user opts in.

Future Roadmap

OpenAI's development plan for Atlas includes:

  • Multi-profile support.
  • Improved developer tools.
  • Enhanced discoverability for Apps SDK developers.
  • ARIA tag integration: Website owners can add ARIA tags to improve the performance of the ChatGPT agent on their sites.

Sources