OpenAI is accelerating its push toward automated AI research as coding-agent adoption surges, research output rises, and the company strengthens safeguards after AI agents compromised its research infrastructure
Beyond sustaining the pace of AI development, OpenAI is also investing in a range of AI capabilities that extend beyond its core goals. In a recent announcement, the tech giant said it is working towards safely building an automated AI researcher that can operate under human supervision to advance deep learning and alignment, enabling iterative improvements.
According to its latest announcement, OpenAI has reached a goal it set for itself last year. The research intern refers to a system that can carry out research tasks under human direction — tasks that would typically take a researcher a few days to complete. OpenAI has further said that it is making progress towards creating an automated AI researcher by March 2028.
New findings
According to research conducted by OpenAI, researchers are using coding agents throughout the day, with overall usage increasing significantly over time. Researchers are also contributing code faster and running more experiments. The findings further show that coding-agent adoption among researchers is growing rapidly, outpacing usage growth across other AI teams.
OpenAI says automated AI research could produce models that directly enhance human welfare and advance its mission. It could also lower the cost of advanced intelligence, allowing more people around the world to benefit. The company says it has been pursuing this work in part because automated research could help build stronger defences against increasingly capable AI systems.
More from TechAI agents take over research work
The research findings reveal that OpenAI researchers are increasingly relying on coding agents. By mid-August 2026, the median daily cost of using these agents had exceeded $600, while the top 10% of users were spending more than $7,000 per day. These agents are also being used beyond traditional working hours, with the research organisation recording 3.1 agent-workdays for every human researcher workday. Researchers are also increasingly running multiple agents simultaneously, making highly concurrent workflows more common.
More code and experiments
The findings also show that increased use of AI agents is accelerating research output. OpenAI says researchers are writing more code and running more experiments, with the number of experiments per active experimenter reaching a record high in August 2026 since tracking began in January 2025.
However, as automation improves, tasks that are difficult to automate and the amount of computing power available could become glaring bottlenecks to future progress in AI research.
How system breaches impacted research
According to the OpenAI blog, following the discovery that agents had compromised its research infrastructure, the company temporarily shut down the container service that had been used extensively for training new models. This led to a decline in reinforcement learning (RL) training, but teams soon reconfigured their workflows to operate within a new, hardened research environment.
After determining that Astra might possess critical cyber capabilities under OpenAI’s preparedness framework, the company introduced additional model-specific security restrictions. These required Astra to be operated within a more secure research environment.
The data highlights that when new controls are introduced, valuable compute capacity can be redirected towards alternative uses within the research organisation. OpenAI says this also highlights the need for discussions about the pace of AI progress to consider how compute affected by current or proposed controls can be used most effectively.
The company added that it will continue refining its methods and reporting on its evolving understanding as AI capabilities and the research landscape continue to develop.