With ChatGPT Atlas, OpenAI seeks to redefine how users browse the Internet by directly integrating artificial intelligence within the browser itself. ChatGPT Atlas, launched by Open AI CEO Sam Altman, merges traditional navigation with an AI digital assistant capable of understanding what you are doing, helping the user complete tasks, summarize information, write text or even automate complex tasks, without ever having to leave the page they are on. The main functions of the browser that challenges Google and AI Mode revolve around three elements: “browser memory”, which allows AI to remember the context of its searches and interactions; the “agent mode”, which allows ChatGPT to act directly on the Web on your behalf, just like an agentic AI; and an advanced privacy control system, which lets the user decide at any time what the AI can see, remember or forget. Atlas is already available for macOS for users on the Free, Plus, Pro and Go plans, and will soon also be available for Windows and as an app for Android and iPhone.
The features of ChatGPT Atlas: what it can do
Let’s start from the basics of ChatGPT Atlas: once you have downloaded the browser and started it, you need to log in to your ChatGPT account and you can then import bookmarks, passwords and history from the previously used browser. The new home tab becomes a single starting point for typing a web address or asking a question directly to ChatGPT, receiving text, visual, or video results in a single stream. One of the most intuitive features is the “Ask ChatGPT” sidebar, which can be opened on any page to have the AI analyze the content you are reading. This way you can, for example, have an article summarized, clarify a concept or have a text you are writing sorted out, all without changing windows or copying and pasting information.
A distinctive aspect of Atlas is its ability to remember information. So-called “browser memories” allow ChatGPT to retain key details about pages visited, searches performed and topics that interest the user. This allows him to retrieve those contexts when needed: you can ask, for example, “show me the job offers I was looking at last week” and then get an updated summary with industry trends. The browser memory integrates perfectly with that of ChatGPT and, as a result, conversations can draw on past chats and details to help the user complete new tasks. It should be specified that these memories are optional and entirely under the control of the user, who can view, archive or delete them at any time. Deleting the browsing history also automatically deletes the associated memories.
Memory also serves to personalize interactions, allowing the AI to offer suggestions consistent with what you have done in the past, such as resuming an interrupted search or continuing the exploration of a related theme. All stored data remains private and tied to your ChatGPT account. If you want the pages you visit to contribute to the improvement of OpenAI’s language models, you can also voluntarily enable the “Include web browsing” option in the data controls settings; by default, however, no information is used to train OpenAI models.
Another very interesting feature of ChatGPT Atlas is the “agent mode”, which introduces a more dynamic level of interaction. In this mode, ChatGPT does not simply respond to inputs provided by the user, but can perform actions within the browser: open tabs, click buttons, search for data or perform small automations. All this, without the user having to carry out the operations in question. Just to give you an example, imagine you want to book a hotel for a weekend away: you can provide ChatGPT with all the relevant details and your specific preferences and needs, asking it to find the most suitable solution and complete the booking independently. The Agentic AI built into ChatGPT Atlas is currently available in preview for Plus, Pro, and Business users.
The security measures of the OpenAI Agent
OpenAI has implemented a number of measures to ensure security and control. For example, the agent cannot run code in the browser, download files, install extensions, or access other applications on the computer or its file system. If it has to act on sensitive sites, such as banking ones, it stops and asks for confirmation before proceeding. It is also possible to use “disconnected mode”, which limits access to data and mitigates as much as possible the risk of unwanted actions taken by the AI agent. As the developers explain, in fact, agents can still make mistakes or receive malicious instructions hidden in web pages or emails, which is why Atlas includes adaptive security systems that are constantly updated.
