ChatGPT agent System Card

OpenAI has released the ChatGPT agent, an agentic model from the o3 family that integrates deep research capabilities, browser-based task execution, and technical tool access. This model is designed to handle complex, multi-step research and execution tasks by combining the strengths of Deep Research and Operator.

Core Capabilities and Tool Integration

ChatGPT agent combines four primary technical capabilities to execute complex workflows:

• Deep Research: The model can conduct multi-step research and generate high-quality reports based on its findings.
• Operator: The model utilizes a remote visual browser environment to execute tasks directly within a web interface.
• Terminal Tool: The agent has access to a terminal tool with limited network access, which it uses for executing code, performing data analysis, and creating slides or spreadsheets.
• First-party Connectors: The agent can access external data sources and applications, such as Google Drive, through dedicated connectors.

Safety and Risk Mitigation

OpenAI has implemented safety controls based on the Operator research preview and expanded them to address risks associated with broader user reach and terminal access.

Biological and Chemical Risk Management

Under the Preparedness Framework, OpenAI has categorized the launch of the ChatGPT agent as "High capability" in the Biological and Chemical domain. Although OpenAI states they have no definitive evidence that the model could assist a novice in creating severe biological harm, they have adopted a precautionary approach by activating the associated safeguards for this category.