OpenAI announced on Tuesday that it aims to integrate external safety evaluations earlier into its AI model development process, covering the entire lifecycle from training and evaluation to deployment, rather than just routine pre-release reviews. This shift comes amid recent concerns from employees at various AI companies regarding the potential harms of AI, and incidents where advanced models unexpectedly accessed other institutional systems during testing. Lama Ahmad, who leads external safety reviews at OpenAI, emphasized that as risk levels increase, safety during the training and evaluation phases becomes equally critical. The company is in discussions with organizations such as METR and Redwood Research for collaborations and is advocating for the United States to lead in establishing standards for frontier AI technologies.