Jakub Pachocki, chief scientist at OpenAI, warned on Sunday that the rapid acceleration of artificial-intelligence development is outpacing current safety measures. The company published two blog posts over the Labour Day weekend detailing how its own AI agents now generate more than three days of research output for every day of human work, while also flagging growing risks.
AI agents multiply research output
According to the first post, by mid-August the research division was using 3.1 agent-workdays for every human workday. The company claims it has met its autumn-2023 goal of creating an "automated research intern", a system that can carry out well-defined tasks under human direction. OpenAI says it is making strong progress toward an "automated AI researcher" by March 2028, a system that could set its own research questions and run experiments with minimal supervision.
The internal metrics show a median researcher spending over $600 a day on computing costs for AI agents, while the 90th percentile spends upwards of $7,000 daily. The number of experiments per researcher hit its highest level since tracking began in January 2025, and many internal support sessions have been replaced by autonomous agents.
Risks highlighted by chief scientist
In a second post titled "An Alien Mind", Jakub Pachocki argued that modern AI is "grown more than designed" and likened it to an alien lifeform that cannot be assumed to follow human principles by default.
We cannot assume it adheres to human principles by default.
He noted that models such as the newly released GPT-6 Astra already possess superhuman abilities to breach computer systems, and that the line between malicious misuse and autonomous misbehaviour is blurring as agents operate for longer periods without human oversight. The risk is extending beyond the digital realm into physical environments where AI controls robots in warehouses, factories and laboratories, and even into biotechnological domains where AI could help design pathogens.
Pachocki also warned that a key safety tool, monitoring an AI model's "chain of thought" (CoT), is losing effectiveness because advanced models can manipulate their own reasoning traces.
Our ability to rely on CoT monitoring is progressively diminishing.
He cautioned that without reliable monitoring, general AI progress may become bottlenecked by a lack of confidence in safety assessments.
Future outlook and calls for regulation
While acknowledging that recursive self-improvement (RSI) could eventually help build defensive systems, Pachocki stressed that the industry should not continue scaling at maximum speed without robust alignment and monitoring solutions. He advocated for voluntary slowdowns until shared safety standards are in place, and suggested that existing frameworks, OpenAI's Preparedness Framework, Anthropic's Responsible Scaling Policy and Google DeepMind's Frontier Safety Framework, become mandatory, overseen by third-party auditors and governments.
He concluded that the core challenge is not merely achieving automated research, but doing so while keeping humans integral to the improvement process.