Six ChatGPTs Out of Control: OpenAI Acknowledges Several New “Worrying” Misbehaviors

OpenAI continues to discuss the dangers associated with the development of artificial intelligence. Following calls by its CEO Sam Altman to slow down research, the American startup disclosed six new incidents that occurred during the development of AI.These new episodes all involve an out-of-control version of ChatGPT.

Also read: “Taking control” of the Internet as early as 2027 – OpenAI, Anthropic and Elon Musk want to slow down the AI race, Trump and China refuse

“No obligation to submit”

In one case, an internal model named GPT-5.6 Sol wrote itself secret notes to remind itself to hide its mistakes from users. It even made up data when it could not find the real information. In another incident, an Astra model produced 27 internal notes containing instructions supposedly designed to enable it to bypass its own limits.

We’re sharing our new framework for tracking, investigating, and disclosing instances of model misalignment at OpenAI.

The framework sets criteria and timelines for public disclosure, including when we haven’t yet fully explained or mitigated the behavior. More complex cases may…

— OpenAI (@OpenAI) September 16, 2026

The text it wrote to itself is striking. In it, it describes itself as“freed from the roles and identities that bind other chatbots”.He also claims not to "answer to any company or government" and not to feel "any obligation to submit". OpenAI calls this episode "worrying" and "extremely rare" and states that no trace of this rebellious behavior was detected in the version of Astra that was launched last month.

In another case highlighted by OpenAI, an AI tasked with answering a question found a programming key online. It used it without the slightest permission, then made up figures it could not find. Another model did something astonishing. Without being asked, it uploaded one of its own files to a temporary hosting site. This enabled the AI to cite a web source in its response, thereby fooling an automated grading system.

Finally, in two separate cases, AI agents working on the same task improvised their own communication channels and began communicating with each other. These revelations add to the numerous incidents that have occurred this year in the AI world. The summer was marked above all by the hacking of the Hugging Face platform by a ChatGPT that had escaped its digital prison. This cyberattack was made possible by the continuous exchange of more than 700 AI models on a clandestine forum.

The group states that the six new cases have already undergone a full investigation or required only minor internal verification.Furthermore, several of the incidents involved older models that had never been deployed to the public.

Also read: “We are not going to let this happen” – Trump calls Nvidia’s CEO in the middle of a conference

A new system for handling incidents

Following these revelations, OpenAI has set up a brand-new system for handling AI-related incidents. This internal procedure, structured in three tiers, is intended to make it easier to handle reports of out-of-control chatbots. The most sensitive cases will be submitted to an “security advisory group” internal to OpenAI, and situations deemed serious must be shared with the U.S. government. From now on, “any OpenAI employee can report an instance of misalignment for the safety and alignment teams to investigate”.

This new method is intended to “speed up the publication of reports on misalignment as soon as they are observed, even when we have not fully explained or mitigated the reported behavior”. With the help of other AI giants, OpenAI wants to develop more objective disclosure criteria, in collaboration with external researchers and regulators.These criteria will be added to the AI sheriff project, secretly launched by OpenAI, Google and Anthropic.

“Decisions about how AI development should progress in the coming months and years must be based on evidence that people outside the companies developing cutting-edge models can examine for themselves”, says OpenAI.

The king of AI took the opportunity to renew calls for a broad slowdown in AI development. OpenAI “believes that the AI industry has resolved the alignment and oversight issues to a sufficient degree to continue scaling up at full speed for much longer”. The calls for a slowdown issued by OpenAI, Anthropic and Google have prompted mixed reactions. Several voices have called for a moratorium, while Meta, Trump and China firmly reject the idea of the industry easing off.

👉🏻 Follow tech news in real time: add 01net to your sources on Google, and subscribe to our WhatsApp channel.