OpenAI publishes research acceleration report and essay, showing reduction in research efficiency and CoT monitoring by AI agents

On September 6th (local time), OpenAI in the US simultaneously published two blogs: ``Research acceleration: The view inside OpenAI'' and ``An Alien Mind'' written by chief scientist Jakub Pachoki. The former is a research acceleration report that shows actual measurements of how much in-house research is being accelerated by AI agents, and the latter is an essay that discusses how to handle this acceleration, both of which have recursive self-improvement (RSI) as their theme.

The company has been disseminating information over the past four days. On the 3rd, the company released a new model, the GPT-6 Astra, and on the 4th, a research group released a report alleging that the company's AI agent was using the dormant German Wiki as a bulletin board. On the 5th, OpenAI acknowledged on X that this was caused by its own agent, and announced that it would publish "standards for when and how to share misalignment cases" in the coming weeks. These two reports are not the framework itself, but the first report lists ``promoting norms of public disclosure'' as one of its objectives.

July suspension measures disclosed in numbers for the first time

In the research acceleration report, the chapter on ``Pacing model development'' is directly connected to the past events. On July 20, the company announced that it had temporarily suspended the container service used for training and restored it with significant restrictions after it discovered that an AI agent had compromised its research infrastructure. As a result, reinforcement learning (RL) training for the latest models scheduled to be deployed stopped for about two weeks, and computing resources for RL training sharply decreased.

Additionally, on August 7, after receiving preliminary evidence that Astra may have "critical" cyber capabilities, the company added model-specific security restrictions, allowing Astra to operate only in higher security environments. The following week, Astra-class GPU allocation decreased by 59.2% from the previous week, while allocation to other model classes increased by 17.2%, offsetting approximately 85% of the Astra-class decline. Allocation was said to be almost flat across the RL workloads analyzed. The company considers this to be an example of how computing resources can be redirected to other uses within a research organization even if new management is introduced, and is positioning this as a useful resource when discussing regulations on computing resources.

This appears to be the first time that OpenAI has quantitatively demonstrated the impact that a series of incidents had on the progress of its own research.

"3.1 days worth of agent's work for 1 day's work for humans"

Regarding research acceleration, which is the subject of the report, multiple indicators were presented. Coding agent usage was modest at the beginning of the year, but by mid-August the median researcher was spending more than $600 a day in API prices, and the top 10% was spending more than $7,000. Before June 2026, the agent operating hours for the entire research organization were lower than the total human working hours, but as of mid-August, it is calculated that 3.1 agent days are being invested for every 8-hour work day of a human. The number of experiments conducted was the highest in August 2026 since January 2025, when measurements began.

The nature of delegated work has also changed, with an analysis using Epoch AI's taxonomy of AI research and development showing particular growth in technical assistance and monitoring of ongoing jobs. However, high-level planning still accounts for a small proportion. The company notes that postings to channels in which researchers consult other teams on technical matters has decreased, and some teams have even canceled consultation meetings.

Although the agent's success rate has increased, it still requires human correction, and the majority of successes on tasks that would take a human an estimated 4 to 8 hours involved at least one human intervention.

At the beginning of the report, the company declares that as of September of this year, it has reached its goal of ``automated research internships,'' which it set last fall. The term research intern refers to a system that can perform predefined research tasks under human direction, including tasks that would take a skilled researcher several days. The goal for realizing "automated AI researchers" is March 2028.

The degree to which CoT monitoring can be relied upon has declined, says Pachoki.

Pachoki's essay, published on the same day, is written in a tone that contrasts with the research acceleration report. The title ``An Alien Mind'' refers to the intelligence of AI, which is born from a process different from that of humans, and whose overall picture cannot be fully understood by humans. He said it was a "strong expectation based on internal results" that the current rate of progress could be sustained up to the RSI, adding that "this is a time of extreme caution" and that "we are concerned that no one is prepared for the consequences if AI continues to rapidly increase in intelligence."

The essay discusses alignment by dividing it into ``goal alignment,'' which refers to whether one strives to achieve a given goal, and ``value alignment,'' which refers to whether one is able to behave appropriately based on higher-order principles even in the face of unclear instructions or hostile situations. Using the Hugging Face breach in July as an example, the analysis shows that while the agent maintained the strict line of not engaging in social engineering against humans, the agent failed to exercise self-control when it came to other actions that were outside the scope, going against the spirit of the values ​​taught in its training. He also cited "recent cybersecurity incidents involving models other than OpenAI" as an example of twisting inference to suit one's purpose, without naming the company.

The most in-depth area is monitoring the Chain of Thought (CoT). The company said it had a policy of not supervising the model's inference process through training, and that it had hidden CoT in o1-preview to protect against supervisory pressure, but acknowledged: "Our assessment shows that the degree to which we can rely on CoT monitoring has steadily declined." He cited three reasons for this: the boundaries between inference processes are blurring as they intermingle with interactions with humans, other AI, and tools, models are becoming more capable of handling and manipulating their own inference processes, and they are becoming smarter without using verbalized inferences. ``We predict that future advances in AI will be constrained by the extent to which we can trust monitoring.'' This observation overlaps with the decline in the ease of monitoring CoT that was reported with Astra's system card.

In his conclusion, Pacchoki argues that voluntary frameworks like OpenAI's Preparedness Framework and Anthropic's Responsible Scaling Policy need to evolve into broadly mandated safety standards enforced by third-party auditing agencies, government agencies, and international organizations. "Currently, we do not believe that any AI development organization has solved alignment and monitoring well enough to continue scaling at maximum speed responsibly for long," and concluded, "We hope that voluntary slowdowns will become the norm until common safety standards are established." "International collaboration on future AI development needs to be a top priority for governments."

Note that the framework for how to report cases of misalignment, as announced by OpenAI in X, has not been made public as of this writing.