The company OpenAI announced an update to the model that operates in the Operator service — an AI-based agent that can autonomously browse web pages and perform tasks in a virtual environment. Soon, Operator will transition to a model based on o3, which belongs to a new series of models for solving logical and mathematical problems. Previously, the service used a special version of GPT-4o.
The new version is named o3 Operator and has been configured with computer safety in mind. For this, the company used special datasets that help the model make decisions about confirming or denying certain actions.
In the technical report, OpenAI noted that o3 Operator is better at protecting against so-called instruction-based attacks in user queries and less frequently performs unwanted or potentially dangerous tasks compared to the previous model. The model employs a multi-level security approach similar to that used in the GPT-4o-based version. At the same time, despite improved code handling capabilities, o3 Operator does not have direct access to the software environment or terminal.
