OpenAI's internal AI model is now capable of autonomously taking over the entire training process for experimental AI models. Multiple agents can spontaneously collaborate, significantly reducing experimentation time. Additionally, they can independently write programs to optimize GPU kernels, representing the initial form of RSI. OpenAI has released a global safety initiative, noting that while fully autonomous RSI has not yet occurred, its development is accelerating. Without proper controls, humanity may lose practical control over AI development. To address this, OpenAI calls for the establishment of globally unified technical standards for frontier AI models, defining rules such as assessment of AI automation in research and development proportions, mandatory human supervision triggering mechanisms, and incident classification reporting. It also proposes that the United States lead this process to ensure that safety alignment research progresses faster than AI capabilities. Furthermore, OpenAI is finalizing an agreement with Anthropic, which previously left due to differences in safety philosophies, to conduct cross-testing of each other's commercial models. This mutual inspection aims to avoid blind spots in self-audits and address the unknown risks posed by the approaching RSI.