Do you want smarter insights in your inbox? Sign up for our weekly newsletters to get what is crucial for AI leaders, data and security. Subscribe now
Openai does not allow Delay in your AI Open Source model Release from sending other functions.
Today The company will reveal Agent chatgptA function that allows AI chatbot autonomous network viewing, conducting extensive research, downloading and creating up-to-date files for your human users using your own virtual computer.
Return? Chatgpt gets it now own PC? And can he employ this computer to log in to your human user, account and download or send things for you?
This is true, at least in a virtual sense, according to Opeli. As the company explains:
The AI Impact series returns to San Francisco – August 5
The next AI phase is here – are you ready? Join the leaders from Block, GSK and SAP to see the exclusive look at how autonomous agents transform the flows of the work of the company-decision-making in real time for comprehensive automation.
Secure your place now – the space is circumscribed: https://bit.ly/3guplf
“The model can open the page using a text browser or visual browser, download the file from the Internet, manipulate it by launching the command in the terminal, and then display the output data back in the visual browser. The model adapts its approach to performing tasks at speed, accuracy and performance.”
How to employ a chatgpt agent
Users can engage the agent by clicking the “Tools” button in the CHATGPT monitor field, opening the developed menu and selecting “Agent mode” from the available options.
Then, when it is turned on, he describes the task in a uncomplicated language, and the agent can carry it out in internet environments and local applications, combining reasoning with activities that only human user can do manually on his own machine.
Agent ChatgPT can connect to applications such as your personal or business Gmail and Github so that it can download useful information – E -Mile or code – from the account to facilitate with the tasks you are asking for. It can connect to interfaces of programming external applications (API) to download information and employ connected applications and services.
If a website needs logging in, you can do it safely using a special browser view that allows an agent to dig deeper and handle more personalized tasks, such as checking the inbox or completing the forms on your behalf.
Going where the operator could not – offline
The up-to-date CHATGPT agent is based on the Openai “Operator” agent issued in January 2025, which allowed CHATGPT to browse the network and complete forms, place orders and perform other internet tasks in a private “headless browser”, i.e. a cloud -based non -standard web browser, which Openai maintained and offered for every operator session.
However, the operator was circumscribed only to interaction with websites and web applications-not programs that could also work locally on a computer, such as tabules of spreadsheets and software for presenting the deck of slides.
The up-to-date CHATGPT agent can view websites, interact with online forms, run the code, analyze data and provide ready outputs – such as editable presentations or spreadsheets – based solely on user manuals.
The exposure appears on the heels of the report published days ago by the independent website of the subscription technology industry Information suggesting that Openai will update chatgptt Be a more direct competitor of your own investor Microsoft Office Applivers (e.g. Excel, Word, PowerPoint, etc.)
Operator connecting and deep research in one agent
In fact, Opeli positions Agent ChatgPT as a combination of two of its previous agents-operator and deep research, which was introduced in February 2025, which searches exhaustively with the facilitate of its own browser exclusively a textual browser to find and develop information in long-term and free reports (hence the name). As openai writes in the blog post:
“The operator could not delve into analysis or write detailed reports, and deep research could not interact with websites to improve the results or access the content requiring user authentication. In fact, we have seen that many questions that they tried with the operator were better suitable for deep research, so we collected the best of both.”
The operator’s previous tool will be withdrawn, but users can still access deep research by developing a menu in the chatgpt interface.
Regardless of whether he uses a visual browser to interact with the website or the Python code starting terminal, the agent smoothly moves between the tools in one session.
It supports a wide range of employ, from competitors’ analysis and generation of reports to travel planning, summary of E -Maili or booking meetings.
Users can interrupt, redirect or stop the task at any time, and the agent receives where it ended.
Availability and access
From today, subscribers of the “PRO” CHATGPT “PRO” level in the amount of USD 200 will have full access to the ChatgPT agent, with a monthly amount of 400 messages.
Chatgpt Plus (USD 20 per month) and a team (30 USD per month) will gain access in the next few days, with 40 messages per month. Additional employ is available using credit options.
Opeli said in the VentureBeat version as part of the embargo that his subscribers ChatgPT Enterprise and Education will gain access to functions in the coming weeks.
For now, the function is not yet available in Europe or Switzerland, undoubtedly disappointing residents there.
Built with safety and control at the forefront
Considering that the agent can now take action on behalf of users-on this logged-on websites or with connected applications-openai, he has introduced extensive security measures.
They include confirmation of users before taking action, energetic supervision over sensitive tasks and technical security to reduce unintentional behavior.
Key security includes:
- Confirmation hints Before actions such as sending forms or sending E -Maili
- Observation modewhich stops the performance when the user becomes inactive
- Refusal of high -risk taskssuch as financial transfers or privacy violations
- No memory retention During the agent session
High -risk domain classification
According to the ready -made frames, Opennai treats the chatgpt agent as High ability System in biological and chemical domains.
Although there is no direct evidence of improper employ, the company activates its strongest security security.
They include improved refusal training, red teams by bios experts and improved detection systems.
Remember that Anthropiku has published the latest information about the up-to-date Claude Opus 4 and Other surveys of advanced AI modelsThey showed that by receiving access to external tools and applications such as E -Mail, in some cases they can take actions that they consider moral and ethical, but they can threaten the user, such as sending E -Maili to government agencies or journalists suspected of user offenses.
The model believes that it behaves like a “guide”, but in fact he can threaten the privacy of users, security and reserved information, as well as warning the authorities about the offense where there is no or where it is doubtful.
Steep performance on tasks in the real world
Agent Chatgpt is not more capable of theoretically-he has filled robust results in many comparative tests designed to simulate real knowledge of real knowledge. He determined a up-to-date high result at the last examination of humanity with 44.4 using parallel implementation methods and reached 27.4% in arduous reference to Frontiermath.
He obtained 45.5% in the spreadsheet in the spreadsheet – more than doubles Copilot in Excel performance.
Current restrictions and subsequent steps
Some functions, such as slide generation, are still beta and may feel basic in formatting or slightly differ between the previews in the application and exported files.
Opeli actively trains another iteration of this function to improve the Polish and system.
Launching the CHATGPT agent signals a change in the interaction of users from AI – from asking questions to assigning full tasks.
Thanks to the ability to reason, act and produce results, OpenAi bet that users are increasingly wanting AI to facilitate them, but also for them. While the company emphasizes that the agent is still developing, it perceives this premiere as the basis for a more interactive, future -oriented artificial intelligence.
