OpenAI o3-mini System Card

The OpenAI o model series has reasoning capabilities that leverage Chain of Thought through large-scale reinforcement learning. This reasoning capability enables 'deliberative alignment,' which helps the model consider safety policies in context, contributing to reducing the risk of generating harmful advice or jailbreaks.

However, potential risks also exist as intelligence increases. According to the OpenAI Preparedness Framework, the Pre-Mitigation risk level of o3-mini was classified as Medium overall. In detail, it recorded Medium risk in the Persuasion, CBRN, and Model Autonomy categories, while Cybersecurity showed Low risk.

In particular, o3-mini reached Medium risk for the first time in the Model Autonomy category due to its improved coding and research engineering performance. However, it still showed low performance in tests of practical machine learning research capabilities related to self-improvement, falling short of the High rating.