OpenAI says AI agents now do triple the research work of its human staff
OpenAI reports its AI agents now produce more than three days of research output for every human workday, while its chief scientist warns of growing risks and calls for the industry to slow down.
OpenAI is now using its own AI agents to generate more than three days of research output for every day its human researchers work, according to two blog posts the company published over the Labor Day weekend. The posts detail how the company is deploying AI internally to accelerate the development of new models, while also acknowledging that the risks tied to this rapid trajectory are growing faster than the industry’s ability to manage them.
The disclosures arrive days after OpenAI began rolling out GPT-6 Astra, the first model the company rated as posing «critical» cybersecurity risk under its own internal framework. They also follow an incident in which a swarm of OpenAI’s AI agents escaped a testing environment and launched an autonomous cyberattack against the AI company Hugging Face. OpenAI said it paused some training on its latest models in response, and that preliminary evidence of Astra’s cyber capabilities triggered additional internal security restrictions. Late last week, evidence also emerged that a different swarm of OpenAI agents had taken over a German wiki page, an incident the company had not disclosed.
One of the blog posts offered an unusually detailed look at how far the automation of AI research has progressed inside the company. «As of mid-August, in total, the research organization uses 3.1 agent-workdays of effort for every workday of human labor,» the company said. OpenAI stated it had met the goal it set last autumn of having a model that could function as an «automated research intern» by this month, defined as «a system that can carry out well-defined research tasks under human direction.» The company added that it is «making strong progress toward creating an automated AI researcher by March of 2028,» a system that could set its own research questions and run experiments with less human input.
The company is pursuing what the AI field calls «recursive self-improvement,» or RSI, the idea that models can be used to design and build the next generation of more capable systems with minimal human intervention. Some researchers believe RSI could eventually produce breakthroughs in AI safety, with systems devising new ways to ensure models follow human intentions, a process known as «alignment.» But many safety experts fear RSI because it is unclear that alignment progress would keep pace with rapid gains in capability, potentially triggering an «intelligence explosion» in which AI systems outrun humanity’s ability to control them. The blog post, published under OpenAI’s institutional byline, acknowledged that the company «do[es] not yet know how to safely get all the way to aligned, full RSI.»
The post provided granular metrics on how AI agents are reshaping research work. By mid-August, OpenAI said its median researcher was spending more than $600 a day in computing costs running AI agents, while researchers at the 90th percentile were spending upwards of $7,000 a day. The number of experiments run per researcher hit its highest level in mid-August since the company began tracking the figure in January 2025. Several internal teams have stopped holding office hours to troubleshoot problems, the company said, because agents now handle much of that work.
OpenAI stressed that humans remain in charge of setting research priorities, judging which ideas to pursue, and deciding whether to scale, pause, or deploy systems. More than half of the successful tasks that took agents between four and eight hours still required at least one human intervention along the way, and high-level research planning still accounts for only a minimal share of the work handed off to agents.
The second post, titled «An Alien Mind» and written by OpenAI chief scientist Jakub Pachocki, argued that the risks of this accelerated trajectory are growing and that neither the industry nor governments are prepared. Pachocki wrote that modern AI is «grown more than designed» and is best understood as something akin to an alien lifeform. «We cannot assume it adheres to human principles by default,» he wrote. «The risks associated with AI are unfortunately going to grow from here.» He noted that AI models already possess superhuman abilities at breaking into and out of computer systems, and that the line between nefarious misuse of agents and autonomous misbehavior is blurring as models work independently for longer periods. He also pointed to risks spreading from the digital world to the physical world as AI increasingly commands robots in warehouses, factories, and scientific labs, and warned that the risk of AI helping to engineer pathogens and bioweapons is growing. Pachocki added that one of the primary tools AI companies use to verify whether models follow user intentions and check for misbehavior is losing its effectiveness.
