Chatgpt has learned to perform tasks instead of a person

Openai presented a new universal SI agent in Chatgpt, capable of performing a wide range of computer problems on behalf of the user.

Chatgpt Can Now Do Work for You Using ITS OWN Computer.

Introduction Chatgpt Agent-A unified agentic system combining Operator’s Action-Taking Remote Browser, Deep Research’s Web Synthesis, and Catgpt’s Conversational Strenghths. pic.twitter.com/7un2nc6nbq

– Openai (@openai) July 17, 2025

According to the company, the agent can automatically manage the calendar, create edited presentations and slides, and run the code.

Catgpt Agent combines the functionality of several previous agent systems: Operator clicking interact with websites, the possibilities of Deep Research to collect information from dozens of sources and create concise analytical reports.

The agent is managed in the dialogue in the usual language. Initially, the tool is available to users with Pro, Plus and Team. To activate, you need to select Agent Mode in the Chatgpt menu.

The new agent supports Catgpt Connectors – integration with Gmail, GitHub and other applications for requests. He also has access to the terminal and is able to work with API.

Use scenarios already include planning and buying ingredients for a four breakfast for four, analysis of competitors and creating a presentation based on results.

Testing The model underlying the agent shows the advanced results in a number of benchmarks. In the Humanity’s Last Exam test, one of the most difficult, which includes thousands of questions in more than 100 disciplines-Chatgpt Agent gained 41.6%, twice the results of O3 and O4-mini models.

Comparison of models in Humanity’s Last Exam. Source: Openai.

In a complex mathematical Frontiermath benchmark, the agent has achieved 27.4%. The previous record belonged to O4-mini-6.3%.

Security Openai emphasized the importance of the safety of a new agent because of its powerful functionality, which can potentially be used by malefactors.

The system report states that the model has a “high ability” in the field of biological and chemical weapons, that is, theoretically able to strengthen the existing ways of task. There is no direct evidence of the threat, but the company takes precautions:

• Online monitoring-all user requests pass through a classifier that determines whether they are related to biological topics. If so – the answer is checked again for a potential threat;
• Disabled memory function – to prevent data leakage due to harmful injection in industries.

We will remind, in July it became known that Openai increased the safety of systems against the background of threats of corporate espionage by Chinese competitors.

The gun