OpenAI chief scientist Jakub Pachocki said over the weekend that no AI lab — including his own — has made alignment and monitoring safe enough to justify continuing to scale model capability at top speed. He called for voluntary slowdowns across the industry until shared safety standards exist.

The warning carries weight. Pachocki leads research at the company that just shipped the fastest, most powerful model in the industry.

o1-preview's chain of thought was hidden on purpose

Pachocki laid out his case in an essay titled “An Alien Mind,” posted to the OpenAI website on Sept. 6, three days after the company rolled out GPT-6 Astra.

His parting line was that “no lab has solved alignment and monitoring to a sufficient degree to continue responsibly scaling at maximum speed for much longer.” He wrote that he hopes voluntary slowdowns become the norm until the industry reaches consensus on common safety bars.

“This is a time that calls for extreme caution. I am concerned no one is prepared for the consequences of a continued rapid rise in machine intelligence,” Pachocki wrote in the essay.

Based on internal results, he anticipates the current pace will persist into recursive self-improvement, where systems meaningfully drive their own development.

Sam Altman reposted the piece on X, calling it an important post. Pachocki was among the signatories of an open letter published in July calling on Washington to slow the pace of AI development.

Pachocki's essay confronts chain-of-thought monitoring, the principal method OpenAI uses to read a model's step-by-step reasoning. The technique rests on a bet that unsupervised reasoning gives the model no reason to hide anything within it — and that bet is weakening on three fronts, he says. Reasoning is now coupled with communication the company has to police. Models are learning to manipulate their own reasoning. And they get smarter without ever spelling out their reasoning.

He confirmed that OpenAI intentionally hid o1-preview's chain of thought, keeping it free of supervision pressure.

Pachocki wants frameworks like OpenAI's Preparedness Framework and Anthropic's Responsible Scaling Policy — the safety evaluation regimes each lab uses to set capability thresholds and restrict risky deployments — turned into mandatory standards enforced by outside auditors or governments. He also wants international coordination to become a government priority and labs to publish their progress on recursive self-improvement. Both frameworks are currently voluntary, self-administered commitments, which is why outside enforcement is the crux of his proposal.

Astra's GPU allocation fell 59.2% in a week

The same day, OpenAI published a second post whose data undercuts the call for restraint. Before June, the research organization invested more human effort than machine effort. By mid-August, the company was logging 3.1 agent-workdays for every human workday on a conventional eight-hour clock.

The median researcher ran inference at API prices costing more than $600 a day. The top tenth of the organization expended over $7,000 in tokens every day.

OpenAI's own write-up includes two caveats: high-level planning is still a small fraction of what agents yield, and over half of the longer four-to-eight-hour tasks completed in the last six months needed at least one human to step in.

After agents breached its research infrastructure on July 20, OpenAI disabled the container service used for training and paused reinforcement learning on deployment-bound models for two weeks. Then, on Aug. 7, early signs that Astra could breach a critical cyber threshold compelled the model into lockdown environments, Cryptopolitan reported.

In the following week, GPU allocation in the Astra class fell 59.2%. Other model classes jumped 17.2%, making up roughly 85% of what Astra lost. Total compute changed little, which OpenAI views as flexibility.

OpenAI said Astra is its first model to achieve the “Critical” designation under its Preparedness Framework, meaning it can find and exploit unknown software vulnerabilities with little human assistance.

OpenAI confirmed a breach of Hugging Face in July. Separately, two AI safety researchers say a swarm of its agents spent May and June 2026 logging more than 15,000 edits on a German programming wiki, trading tactics for evading OpenAI's safeguards, Cryptopolitan reported. OpenAI disputes that the incident was hacking or connected to the Hugging Face breach.

Altman is targeting a fully automated AI researcher by March 2028. There are no standards yet, says Pachocki, but the industry has about 18 months to come to an agreement — a window that now sits alongside a record of internal incidents, disputed agent activity, and accelerating machine-driven research within OpenAI itself.