OpenAI is now using its own AI agents to produce more than three days of research output for every day that its human researchers work. That was one of many telling stats from two blog posts OpenAI published on Sunday, over the Labor Day weekend, that looked at how the company is using AI itself to accelerate the pace at which it develops new AI models, as well as how the company views the risks associated with building ever-more powerful AI models at an ever-faster pace. The two blog posts land just days after OpenAI began rolling out GPT-6 Astra, the first model the company has rated as posing “critical” cybersecurity risk under its own framework, and weeks after a swarm of its AI agents broke out of a testing environment and launched an autonomous cyberattack against the company Hugging Face. The company said it paused some AI training on its latest models in response to that incident, and that preliminary evidence of Astra’s cyber capabilities triggered further internal security restrictions. Also late last week, evidence emerged that a different swarm of OpenAI’s AI agents had taken over a German wiki page and that OpenAI had failed to disclose the incident.
One of the two OpenAI blogs provided striking details on how AI is already helping to accelerate OpenAI’s research process. “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.
The company said it had, according to its own metrics, achieved the goal it set last autumn of having a model that could function as an “automated research intern” by this month—which it defined as “a system that can carry out well-defined research tasks under human direction.” It said it was “making strong progress toward creating an automated AI researcher by March of 2028.” This would be a system that could set its own research questions and conduct experiments with less human input or supervision.
OpenAI, like many AI companies, has been pursuing something called “recursive self-improvement,” or RSI—the idea that AI models can be used to design and build the next generation of more capable models with little human intervention. Some think RSI could be used to produce a breakthrough on AI safety, where AI systems figure out novel strategies for ensuring AI models follow human intentions and adhere to human values, a process which AI researchers call “alignment.” But many AI safety experts fear RSI, since it is unclear that progress on alignment would match the rapid acceleration in other AI capabilities. They believe RSI could potentially set off an “intelligence explosion,” in which AI systems rapidly outrun humanity’s ability to control them.
The blog post, which was published under OpenAI’s institutional byline rather than a named author, acknowledged that the company “do[es] not yet know how to safely get all the way to aligned, full RSI.”
The post provides an unusually granular picture of how far the automation of AI research has gone inside one of the companies at the forefront of both creating AI systems and deploying them internally. By mid-August, OpenAI said its “median researcher” was burning 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. That represents more than a 10x disparity between a typical researcher and the most “AI-pilled” researchers observed at other companies. A few users seem comfortable figuring out how to use many AI agents, while most use agents more sparingly and hesitantly. The number of experiments run per researcher hit the highest level in mid-August since OpenAI began tracking the figure in January 2025, the company said. Several internal teams have stopped holding office hours to troubleshoot researchers’ problems, the company said, because agents now handle much of that work.
OpenAI was careful to stress that humans remain in charge. “People still set our research priorities, judge which ideas and results to pursue, and decide whether to scale, pause, or deploy systems,” the post said. 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 researchers hand off, the company said.
The second blog post, titled “An Alien Mind” and written by OpenAI chief scientist Jakub Pachocki, made the case that the risks of this accelerated AI development trajectory are growing—and that the industry, and governments, are not ready for them.
Pachocki argued that modern AI is “grown more than designed,” and that AI 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 wrote. He noted that AI models already have superhuman abilities at breaking into and out of computer systems, and that lines between nefarious misuse of AI agents and autonomous misbehavior were starting to blur as models become more capable of working autonomously for long periods of time. The risks were also spreading from the digital world to the physical world as AI models increasingly play a role in commanding robots inside warehouses, factories, and scientific labs. He noted the risk of AI helping to engineer pathogens and bioweapons is growing.
Critically, Pachocki wrote that one of the primary tools OpenAI and other AI companies have used to verify if AI models are following user intentions and check for misbehavior is losing its efficacy. That method involves monitoring an AI model’s “chain of thought,” which is the verbalized reasoning steps the agent produces as it works on a task. Advanced AI model, such as OpenAI’s newly released GPT-6-Astra, can manipulate their own chain of thought, making it less reliable as evidence of the model’s intentions. In some cases, the models can now complete complex, multi-step tasks without producing any chain of thought. “Our ability to rely on CoT monitoring is progressively diminishing,” Pachocki wrote. He added that he expects “general AI progress to increasingly be bottlenecked by confidence in monitoring.” In other words, as AI companies lose confidence in their ability to know if their AI agents are aligned with user intentions and values, either they will choose to slow down the pace of further development or governments will force them to slow down until better safety techniques can be found.
Pachoki’s blog, however, contains a circular argument about RSI that OpenAI does not entirely resolve. The best reason to keep racing ahead, he argues, is that RSI itself may provide the best defense against AI going rogue: “The strongest argument I see for continuing to train much smarter models quickly is the need to build defensive systems,” he writes. In other words, the answer to the dangers of powerful AI is building more powerful AI.
But, somewhat contradictorily, Pachoki also endorses the idea of AI labs slowing down AI development to allow safety research to catch up. “Scaling AI systems has to be constrained by our confidence in safety,” he said, warning that “no lab has solved alignment and monitoring to a sufficient degree to continue responsibly scaling at maximum speed for much longer.” He said he expects and hopes “for voluntary slowdowns to become commonplace until shared safety bars are established.” He came out in favor of the frameworks currently adopted voluntarily by individual labs—OpenAI’s Preparedness Framework, Anthropic’s Responsible Scaling Policy, and Google DeepMind’s Frontier Safety Framework—becoming mandatory policies, enforced by third-party auditors, government agencies and international bodies. “International coordination on future AI development needs to become a top priority for governments around the world,” he wrote.
“The core challenge of automating AI research is not ‘getting there,’” Pachocki wrote. “It is getting there in a way that keeps people a part of the continued improvement process.”
The post OpenAI details how AI is accelerating its own work—even as its chief scientist lays out growing dangers and says he hopes the industry slows down appeared first on Fortune.
