On September 6, U.S.-based OpenAI announced that it has achieved the goal it set last autumn to realize an "automated research intern" by September 2026. According to OpenAI, a "research intern" is defined as a system capable of performing clearly defined research tasks—which would take a skilled human researcher several days to complete—under human instruction.

The use of AI agents within OpenAI's research division is expanding rapidly. According to data from mid-August, the average user in the research division integrates agents into their daily operations, consuming inference equivalent to over $600 per day in API pricing. In terms of operating hours, it was reported that while agent labor remained below total human working hours before June 2026, the situation reversed by mid-August, with 3.1 days of agent labor being deployed for every one human workday.

Meanwhile, the company admitted that it "still does not know" how to control "Recursive Self-Improvement (RSI)" when AI autonomously enhances its own capabilities. To ensure safety, the company disclosed instances where it implemented pauses when agents intruded into research infrastructure or applied security restrictions to specific models. OpenAI emphasized the importance of ensuring transparency and public discussion so that safety measures can keep pace with the evolution of AI.

Source: