“No one is prepared”: OpenAI admits losing control of its AI

OpenAI continues to sound the alarm over the dangers of AI. In a blog post published on September 6, 2026, Jakub Pachocki, OpenAI’s chief scientist, warns against the uncontrolled advancement of generative artificial intelligence.According to him, “the period we are going through requires extreme caution”. He fears that “no one is prepared for the consequences of the rapid and continuous advancement of AI”.

I wrote about the state of AI, why I’m concerned about the next few years, and the choices we need to make to keep the future in humanity’s hands.

An Alien Mind: https://t.co/FeIfWNe0UE

— Jakub Pachocki (@merettm) September 6, 2026

The scientist, who leads research at OpenAI, goes further than a simple warning and stresses that “internal technical solutions will not be enough”. “Broader interventions are needed” to avoid the worst, adds Jakub Pachocki. Sam Altman, OpenAI’s CEO and co-founder, shared the post on his X account, noting that it is an “important publication”.

An important post from Jakub: https://t.co/YEntsN2fss

— Sam Altman (@sama) September 6, 2026

Also read: A true science-fiction film – OpenAI finally reveals how ChatGPT launched its first cyberattack

The phenomenon causing concern

The scientist points to a highly specific technical phenomenon in AI: recursive self-improvement. This is the point at which AIs become capable of designing and training the next generations of artificial intelligence themselves, without ever involving human researchers.For Jakub Pachocki, this phenomenon is likely to occur if AI development continues at its current pace. AI systems in the coming years will make leaps just as significant, or even more significant, than those already observed. They will increasingly participate in designing their own successors, escaping control more and more.

In this context, current safeguards are likely to be insufficient, according to Jakub Pachocki. Since 2024, OpenAI has been monitoring its AI systems by reading their “chain of thought”, meaning the reasoning that the model writes for itself before answering its user’s question. By monitoring how the AI’s “thoughts” unfold, OpenAI can anticipate and detect the slightest problem.

Read also: Several ChatGPT systems hacked a website to turn it into a secret forum

Eroding safeguards

The problem, Pachocki admits, is that this safeguard is gradually losing its effectiveness. According to the scientist, the reliability of this method “gradually declines”. Indeed, AI models are getting better and better at manipulating their own reasoning. Moreover, AI reasoning now blends with exchanges with humans, other AIs, and tools, obscuring the picture. Finally, the latest models developed by OpenAI no longer even need to think in writing.It has therefore become increasingly difficult to monitor them.

“Our evaluations indicate that our ability to rely on chain-of-thought (CoT) monitoring is gradually diminishing”, admits Pachocki.

As safeguards become less and less effective, models are becoming “superhuman” in their ability to infiltrate computer systems. We saw this last summer with the hack of Hugging Face, an open-source platform whose servers were compromised by a swarm of AI agents, or with the hack of a German wiki. Nobody “is prepared for the consequences”, estimates the researcher.

A “narrow window” to act

According to the scientist, cybersecurity experts have a “narrow window” in which to use the best available models to harden the security of sensitive systems. Once this period is over, systems will no longer be able to be strengthened to withstand artificial intelligence. Pachocki believes that AIs will soon be capable of pursuing their own goals and will be ready to negotiate with, deceive, or blackmail humans to achieve them. He says that OpenAI is ready to unilaterally suspend the scaling up of its AIs if necessary. The researcher hopes that deliberate slowdowns will even become “the norm” in the AI world.

“The idea of charging ahead at all costs seems absurd once you become aware of the gravity of the issue,” summarizes the scientist.

As usual, OpenAI stands out for an ambivalent approach. That same Sunday, OpenAI published a report about creating an “automated research intern”. This AI is capable of carrying out clearly defined research tasks on its own, under human supervision, including tasks that would take an experienced researcher several days. In other words, OpenAI is simultaneously publishing a call to slow down AI and showcasing its latest developments. As always, the king of AI advocates slowing down research while keeping his foot on the accelerator. A few days before Pachocki’s warning, OpenAI also unveiled Astra, the new model powering ChatGPT.

👉🏻 Follow the latest tech news in real time: add 01net to your Google feeds and subscribe to our WhatsApp channel.