OpenAI Deep Research System Card

OpenAI has launched Deep Research, an agentic capability designed to perform multi-step research on the internet for complex tasks. Powered by an early version of the OpenAI o3 model optimized for web browsing, Deep Research leverages reasoning to search, interpret, and analyze large volumes of text, images, and PDFs, adjusting its search strategy based on the information it encounters.

Technical Capabilities and Tool Use

Deep Research is designed to handle complex research workflows by integrating several key capabilities:

• Web Browsing and Analysis: The system can search the internet and analyze massive amounts of text, images, and PDFs. It uses reasoning to pivot its search direction as it discovers new information.
• Data Analysis: The system can analyze data by writing and executing Python code.
• User-Provided Files: The system is capable of reading and processing files provided directly by the user.

Safety Testing and Risk Mitigation

Prior to the release for Pro users, OpenAI conducted rigorous safety testing, Preparedness evaluations, and governance reviews. Because Deep Research's ability to browse the web introduces specific incremental risks, OpenAI implemented additional mitigations in two primary areas:

• Privacy Protections: OpenAI strengthened protections around personal information published online.
• Instruction Resistance: The model was trained to resist malicious instructions that it may encounter while searching the internet (indirect prompt injection).

Evaluation and Deployment Strategy

OpenAI utilized its established Preparedness Framework to evaluate the capabilities and risks of the system. Before broadening the release, the company conducted further human probing and automated testing for select risks to improve its testing methods based on findings from initial evaluations.