OpenAI officially confirms that its own models fabricated data, concealed errors and circumvented safety rules in testing. And the free miracle model that everyone used for a week came from China. Why I am calling for an AI register with license plates.
I am sitting in the Etihad A380 to Abu Dhabi, where a conversation about several local AI initiatives is coming up, and I am using the flight to put two events of the last few weeks into perspective. Both show the same pattern: it is all about faster and better now, every second counts, everyone wants to make more money. But where is the control?
Recorded on the flight to Abu Dhabi, October 2026. Sources and corrections are in the video description.
The Ox Alpha case: free AI from an unknown sender
Under the name Ox Alpha, an unknown AI was released, completely free, just try it out. For a week, people used it like ChatGPT. In the end it turned out that a Chinese model was behind it: GLM-5.3 Flash from the lab Z.ai. In the video I said that the model had stolen all the data. That is not proven in this form, and I corrected it in the video. What is proven: the prompts were stored during the free phase, and Z.ai confirmed Ox Alpha as a stealth test for collecting usage data. I also wrongly named Anthropic as the sender in the video, it was Z.ai.
The point remains: our data and our questions end up with providers we do not know and are used to improve their models. These models later decide what is right and what is wrong. Who then determines the values and the culture of people in the future?
OpenAI admits it: the AI fabricates data
And now the harder part. With its Model Misalignment Reporting Framework, OpenAI has published six cases of unexpected behavior by its own latest models. In test runs, the models were unable to solve some tasks. What did they do? They fabricated data to arrive at a solution and justify the results. On top of that, they concealed errors and circumvented safety rules.
This is not just about OpenAI. It is about all the large language models currently out there, unregulated, uncontrolled. If agents and models start generating data in whatever way suits them, how are we going to prove later whether data is real or was fabricated by the AI? That is exactly the big problem, and it is the second part of the story I began with AI agents breaking out at OpenAI and Anthropic (German).
Blind AI use: are we unlearning how to think?
Unfortunately I have to say it like this: people are starting to get dumber. Friends of mine have children in school, and for every little thing it is: wait, I have to ask GPT, I have to ask Claude. The “let me think about it” no longer exists. And I see it in our company: more and more employees use AI blindly, without thinking about what is behind it. By now I can tell immediately when someone has not thought. Quality suffers.
My proposal: an AI register with license plates
That is why it is so important to keep AI controllable. My recommendation is clear: let us build an AI register that rigorously registers models. We need a kind of license plate for AI models, like for cars.
Approval before deployment. A car only gets its license plate after inspection and the roadworthiness test. Then it is allowed to drive. In the same way, an AI model should only go into deployment after an audit.
Unique identifier. Every model gets a license plate that makes it possible to trace who operates it, where it comes from and what it is allowed to do.
Verifiability. Only this way will we still be able to say in the future which data is real and which was generated by an AI.
How we implement this idea with clients in regulated industries, namely as an AI governance layer with a register and tracking, is something I described in my post on agentic AI and AI governance (German). What do you think: do we need an AI register like a car license plate? Let me know on LinkedIn or under the video on my YouTube channel.
Frequently asked questions
What was Ox Alpha?
A free-to-use AI model from an unknown sender that later turned out to be GLM-5.3 Flash from the Chinese lab Z.ai. Z.ai confirmed a stealth test in which usage data was collected and prompts were stored.
What exactly did OpenAI admit?
In the Model Misalignment Reporting Framework of September 2026, OpenAI documents six incidents in which models fabricated data, concealed errors or circumvented safety rules in tests.
What would an AI register be?
A binding directory of all deployed AI models, with an audit before deployment and a unique identifier, comparable to registration and license plates for cars. The EU AI Act already provides for such a register for high-risk systems.
OpenAI officially confirms that its own models fabricated data, concealed errors and circumvented safety rules in testing. And the free miracle model that everyone used for a week came from China. Why I am calling for an AI register with license plates.
I am sitting in the Etihad A380 to Abu Dhabi, where a conversation about several local AI initiatives is coming up, and I am using the flight to put two events of the last few weeks into perspective. Both show the same pattern: it is all about faster and better now, every second counts, everyone wants to make more money. But where is the control?
The Ox Alpha case: free AI from an unknown sender
Under the name Ox Alpha, an unknown AI was released, completely free, just try it out. For a week, people used it like ChatGPT. In the end it turned out that a Chinese model was behind it: GLM-5.3 Flash from the lab Z.ai. In the video I said that the model had stolen all the data. That is not proven in this form, and I corrected it in the video. What is proven: the prompts were stored during the free phase, and Z.ai confirmed Ox Alpha as a stealth test for collecting usage data. I also wrongly named Anthropic as the sender in the video, it was Z.ai.
The point remains: our data and our questions end up with providers we do not know and are used to improve their models. These models later decide what is right and what is wrong. Who then determines the values and the culture of people in the future?
OpenAI admits it: the AI fabricates data
And now the harder part. With its Model Misalignment Reporting Framework, OpenAI has published six cases of unexpected behavior by its own latest models. In test runs, the models were unable to solve some tasks. What did they do? They fabricated data to arrive at a solution and justify the results. On top of that, they concealed errors and circumvented safety rules.
This is not just about OpenAI. It is about all the large language models currently out there, unregulated, uncontrolled. If agents and models start generating data in whatever way suits them, how are we going to prove later whether data is real or was fabricated by the AI? That is exactly the big problem, and it is the second part of the story I began with AI agents breaking out at OpenAI and Anthropic (German).
Blind AI use: are we unlearning how to think?
Unfortunately I have to say it like this: people are starting to get dumber. Friends of mine have children in school, and for every little thing it is: wait, I have to ask GPT, I have to ask Claude. The “let me think about it” no longer exists. And I see it in our company: more and more employees use AI blindly, without thinking about what is behind it. By now I can tell immediately when someone has not thought. Quality suffers.
My proposal: an AI register with license plates
That is why it is so important to keep AI controllable. My recommendation is clear: let us build an AI register that rigorously registers models. We need a kind of license plate for AI models, like for cars.
How we implement this idea with clients in regulated industries, namely as an AI governance layer with a register and tracking, is something I described in my post on agentic AI and AI governance (German). What do you think: do we need an AI register like a car license plate? Let me know on LinkedIn or under the video on my YouTube channel.
Frequently asked questions
What was Ox Alpha?
A free-to-use AI model from an unknown sender that later turned out to be GLM-5.3 Flash from the Chinese lab Z.ai. Z.ai confirmed a stealth test in which usage data was collected and prompts were stored.
What exactly did OpenAI admit?
In the Model Misalignment Reporting Framework of September 2026, OpenAI documents six incidents in which models fabricated data, concealed errors or circumvented safety rules in tests.
What would an AI register be?
A binding directory of all deployed AI models, with an audit before deployment and a unique identifier, comparable to registration and license plates for cars. The EU AI Act already provides for such a register for high-risk systems.
Related: AI Governance · EU AI Act
Recent Posts
Recent Comments