Safety
Each improvement in model capabilities warrants commensurate improvements to safety. For OpenAI o3 and o4-mini, they completely rebuilt their safety training data, adding new refusal prompts in areas such as biological threats (biorisk), malware generation, and jailbreaks. This refreshed data has led o3 and o4-mini to achieve strong performance on OpenAI's internal refusal benchmarks (e.g., instruction hierarchy, jailbreaks). In addition to strong performance for model refusals, they have also developed system-level mitigations to flag dangerous prompts in frontier risk areas. Similar to their earlier work in image generation, OpenAI trained a reasoning LLM monitor which works from human-written and interpretable safety specifications. When applied to biorisk, this monitor successfully flagged ~99% of conversations in their human red‑teaming campaign.
OpenAI stress tested both models with their most rigorous safety program to date. In accordance with their updated Preparedness Framework, they evaluated o3 and o4-mini across the three tracked capability areas covered by the Framework: biological and chemical, cybersecurity, and AI self-improvement. Based on the results of these evaluations, they have determined that both o3 and o4‑mini remain below the Framework's "High" threshold in all three categories. They have published the detailed results from these evaluations in the system card.