ChatGPT has been taught to perform complex tasks instead of humans

OpenAI has unveiled a new AI agent for ChatGPT, capable of performing a wide range of digital tasks. Users can communicate with it in natural language, and the tool automatically manages calendars, creates slides, analyzes competitors and even makes purchases. It is initially available to Pro, Plus and Team subscribers as part of the agent mode feature.

The new development combines the capabilities of several previous solutions from OpenAI. In particular, the agent can click on websites, collect and summarize information from dozens of sources, run code, connect to services like Gmail, GitHub and use APIs. It also has access to the terminal and ChatGPT connectors.

In the demonstration, the agent handled planning a Japanese breakfast, analyzing competitors, and preparing a presentation. OpenAI emphasizes that the model outperformed all previous versions in a number of benchmarks. For example, in Humanity’s Last Exam the agent scored 41.6%, and in the FrontierMath math test – 27.4%, which is several times higher than previous records.

The company paid special attention to security issues. Since the agent has access to external services and is capable of generating code, protection against potential abuse is provided. All requests are checked, especially in sensitive areas such as biology and chemistry. In addition, the memory function is disabled in this mode.

The OpenAI report categorizes the model as high-risk for bio- and chemical weapons. Although there is no direct evidence of a threat, the company applies preventive measures. This is due to OpenAI’s overall efforts to protect intellectual property and prevent data leaks, including espionage by competitors.