Openai has introduced a big new ability to Chatgpt, which has been launched what it says to the CHATGPT agent. It is a device that can be autonomous complex tasks online, using its own virtual computer. The system combines the interaction of Chatgpt with web interactions and code execution, allowing it to complete the work with minimal human input to complete the work from beginning to end. Openai CEO Sam Altman launched in a post on X, “You can feel AGI”, which is a reference to artificial general intelligence, a long -term goal of making AIs that can logic and learn like a human. “

The new agent mode is now rolling out for Chatgpt Pro, Plus, and Team users in the US, in which the use of enterprise and education is soon expected. Pro users will receive 400 messages per month with new agent features, while other paid users receive 40. Additional use can be added through the credit system. Deep research, one of the integrated devices in the agent, will be available as a standalone option for users that prefer its slow but more detailed results.

With the new agent, users can ask Chatgpt to scan their calendar, briefly, to do things like research competitors and slide decks. It can also shop for materials for a specific recipe or present work expenses. It receives a visual browser, terminal access, text-based web browsing, and API integration with apps such as gmail or zethab using tools.

This agent system makes the earlier equipment of the system, operator, which can interact with the websites, and intensive research, which was good in synthesizing the information. Unlike previous recurrences, which were limited to either browsing or analysis, the new system switchs both liquid between logic and action. It uses its own virtual machine, which helps preserve reference in tasks that require many devices.

Users are given strong control over what the agent does. Chatgpt always asks for permission before doing impressive moves, such as logging in websites, sending emails, or shopping. If necessary, users can stop, even intervene, or prevent tasks. The task can also be determined to repeat, such as generating weekly reports.

Openai insisted that the expanded access to the agent shows new risks. Because it can now work directly on the Internet, including clicking on the link, logging in and handling private data. Therefore, Openai says that it has implemented several security measures. These include the requirement of real -time supervision for certain tasks such as clear user approval for important tasks and sending emails.

One of the greatest concerns is an early injection, where hidden or malicious instructions on a website can cheat AI to behave in unexpected ways. For example, a hidden sign on a webpage may ask the agent to share sensitive information or take harmful action. Openai says that it has trained the model to identify and oppose such attacks, and it monitored for suspicious behavior during tasks.

Privacy is also an important focus. Openai claims that the Chatgpt agent does not store login details, and users can remove all browsing data or log out sites with a single click. Input data during sessions is kept private, and cookies behave according to the individual site policy.

Despite the advances, Openai believes that the agent is still a task in progress. Its ability to make slideshows is currently in beta and can produce basic or inconsistent consequences, especially when it starts from scratches. Slide formatting and exports are being improved, and future updates promise more polished outputs.

While the rollout is currently limited to some users in the US, Openai says that European Economic Zone and Switzerland are still working. Currently we do not know what and when the facility in India will be rolled out. In particular, Openai says that the old operator preview will be closed in the coming weeks.