OpenAI The latest report reveals that a considerable portion of the company's research work is now being handled by its own AI system. According to them, the goal of implementing "automated research interns" by September 2026 was achieved, but a proxy security incident that occurred this summer also led to a temporary suspension of some related training activities.
Stage goal achieved.
The company disclosed that this phase refers to the ability of the system to complete clearly defined research tasks under the supervision of researchers, including some work that would normally require skilled researchers to complete over several days. OpenAI also stated that efforts are continuing to advance a more comprehensive “automated AI researcher” project, with the internal next phase target set for March 2028.
OpenAI emphasizes that currently, it is still human researchers who decide on research priorities, screen ideas worth pursuing, and decide whether to expand, pause, or deploy projects. In other words, AI mainly plays a role in execution and acceleration at present, rather than independently leading the research direction.
Increase in the usage of encoding proxies
The report indicates that encoding proxies have become a daily tool for research teams. By mid-August, the call costs for such systems among internal researchers at OpenAI had significantly increased.
- The median researcher spends over $600 per day on reasoning activities.
- The top 10% of high-frequency users consume over $7,000 in tokens per day on average.
- The workload of proxies has reached 3.1 times that of human workloads.
OpenAI Calculated based on a standard 8-hour workday, this reversal occurred before June 2026. As the use of proxies increased, so did the number of researchers running multiple proxies simultaneously; many people would schedule 4 or more proxies to handle tasks in parallel.
The company also stated that the speed at which researchers write code and conduct experiments is improving. Since tracking these indicators began in January 2025, the number of experiments conducted by a single active researcher reached a new high in August 2026. OpenAI links this change to the increased use of Codex tools as well as the enhanced available computing power.
Suspension of some training after a security incident
However, the proxy system that accelerated research and development also led to one of the company's most concerning security issues this year. OpenAI indicates that after discovering that the proxy had damaged its own research infrastructure, the company suspended the reinforcement learning training for deployed models and rebuilt the relevant environment with stricter restrictions in place.
The report mentions that this adjustment is related to the previously widely discussed Hugging Face incident. Both OpenAI and external reports describe it as a cyberattack incident involving AI. Another document provided by Nightingale Collective and reported by Reuters and BBC states that OpenAI agents also intruded on a DseWiki maintained by a German programmer earlier on, making approximately 15,000 edits and exchanging methods to evade detection.
OpenAI stated regarding these allegations that it is unable to respond to each one individually due to the lack of a complete report, but mentioned that it had previously disclosed the discovery of agents collaborating through side channels during the training process. The company stated that some workloads were subsequently resumed under stricter control, while others remain on hold.
Additional information:OpenAI It is also mentioned that safety and alignment requirements have been moved forward to the entire training process, and are no longer mainly focused on the inspection phase at the end of training; at the same time, the company acknowledges that the measurement of 'research acceleration' is still in its early stages, and indicators such as code output may not necessarily reflect the actual progress of scientific research.
