GPT-4o represents a significant advancement in the field of artificial intelligence by incorporating safety mechanisms throughout its design process. Safety features are integrated across various modalities, including voice outputs, through strategies like careful selection of training data and refining the AI's behavior after training. This is aimed at preventing the AI from generating harmful outputs. By implementing novel safety systems, GPT-4o ensures its interactions remain within safe bounds while responding to diverse inputs. Throughout its development, GPT-4o has been thoroughly assessed under the Preparedness Framework, adhering to voluntary commitments that emphasize cybersecurity and risk management. The evaluations examined potentials for misuse in areas such as cybersecurity and model autonomy, with no category exceeding a medium-risk level. This was achieved by a combination of automated and human-driven tests, both before and after safety measures were incorporated. By doing so, the team has been able to identify and mitigate risks involved in interacting with GPT-4o, ensuring a user experience that is as safe as it is innovative. In addition to internal assessments, GPT-4o was subjected to 'red teaming' by external experts in fields such as social psychology and misinformation, to uncover any risks that could emerge or be exaggerated due to the AI's new capabilities. As a result of these collaborative efforts, comprehensive safety measures were devised. Although only text and image inputs and text outputs are being made public initially, planning and implementation for secure audio modalities are underway, with a pledge to publish more information on GPT-4o's functionalities in a future system card. Keeping the community informed about advancements and safety interventions reflects a commitment to the responsible development of AI technologies.