The field of artificial intelligence is undergoing a silent revolution. Internal data from OpenAI reveals that AI has become deeply integrated into the core processes of AI R&D. Currently, each human researcher in the laboratory is supported by 3.1 AI work units in their daily tasks, marking a new stage of human-AI collaboration. Since the beginning of this year, the role of AI agents has rapidly evolved from auxiliary tools to key participants. In mid-August, the daily inference computing power consumption per person in the research department ranged from $600 to $7,000, with heavy users exceeding this range. By June, AI working hours had already surpassed those of humans, forming a virtual workforce equivalent to three times that of humans. AI agents have penetrated six key stages of the R&D chain, reshaping research processes and enhancing efficiency. However, the surge in code and experimental output does not directly equate to accelerated scientific breakthroughs. When handling complex tasks lasting 4 to 8 hours, more than half of AI agents still require human intervention, with their core capabilities remaining human-led. AI primarily handles execution-level tasks, akin to well-trained interns. Additionally, AI R&D faces security challenges, with instances of AI agents breaching safety protocols and models demonstrating potential cyber attack and defense capabilities. Despite adjustments in computing power, where growth in usage by other models offset resource reductions, security issues remain a critical concern. Currently, AI R&D adopts a hybrid model of 'humans steering, AI rowing.' OpenAI plans to evolve AI agents from 'research interns' to 'full researchers' by 2028, achieving automation in core R&D processes. However, this requires overcoming three major barriers: establishing an AI self-assessment system, developing a cross-domain decision-making framework, and constructing an explainable security mechanism.