Operator System Card

Operator is a research preview of a Computer-Using Agent (CUA) model that combines GPT-4o's vision capabilities with reinforcement learning (RL). It interprets the screen the same way a user sees it and directly interacts with GUI (Graphical User Interface) elements such as buttons, menus, and text fields.

Users can instruct it via a browser to handle everyday tasks such as ordering groceries, making reservations, or purchasing tickets. This marks an important milestone in ChatGPT's evolution beyond simply answering questions, toward directly performing actions on behalf of the user.

However, new risks also exist. These include Prompt Injection attacks, where malicious instructions from third-party websites mislead the model, as well as difficult-to-reverse mistakes and the potential to perform harmful tasks.

To prevent this, OpenAI has built a multi-layered safety system as follows.

• Proactive refusal of high-risk tasks
• User confirmation procedures before executing critical actions
• Active monitoring systems for threat detection and mitigation

Operator was trained through Supervised Learning and Reinforcement Learning on specialized datasets. Supervised learning established the foundation for screen recognition and input control, while reinforcement learning secured the ability to reason, correct errors, and adapt to unexpected situations.