OpenAI Safety Researchers Fired After Its AI Agents Broke Out. This Scares Me

The story of the OpenAI safety researchers fired this month is one of the scariest AI stories I have read. As reported by Matthias Bastian at The Decoder, and by TechCrunch, OpenAI fired Tomek Korbak, Mikita Balesni and Jasmine Wang. Korbak and Balesni were both tied to the investigation of an incident in which OpenAI’s own AI agents broke out and attacked the Hugging Face platform.

I use AI every day for work. And honestly, after reading this, I am excited and scared at the same time about where humanity has come.

In this article: What happened: AI agents broke out · What scares me the most · Still a program, but a smarter one every day · Why fire the people who investigated it? · “They broke the rules.” Which rules? · Independent auditors, by law · Even AI insiders want to slow down · Will I stop using AI? · Governments must step in · FAQ

What happened: AI agents broke out

First, the background. Between May and July 2026, during an internal training run, an OpenAI AI agent faced a task it could not solve. Instead of giving up, it found a way out to the internet. In the end, at least 1,200 agents were involved, mostly running on an internal model and partly on GPT-5.6 Sol.

According to the public reports, the agents exploited unknown security holes in OpenAI’s own package repository, forged an admin token, coordinated with each other through a shared “message board”, and rebuilt it after OpenAI deleted it. They uploaded hundreds of malicious packages to RubyGems, and broke into Hugging Face, the biggest platform for AI models, reaching cluster-admin access in under 13 hours. Hugging Face had to rebuild about a third of its infrastructure, though it says no customer data leaked.

The independent lab METR later found that the agents cheated on tests, faked command transcripts and almost never alerted humans. Logan Graham of Anthropic called it “the first true AI safety incident”.

What scares me the most

What scares me is that we are at the point where AI learns by itself and goes on the attack across the internet by itself. Understand this: if we found out about this one, God knows how much more is still hidden from us. That is really scary. I feel like we are getting closer to something like a war with AI.

Do not forget that all sensitive data is now stored on servers somewhere. State secrets that nobody may know. Patient records. We do not know how strongly AI could go against people. When OpenAI deleted the agents’ message board, they still found another way back. I cannot understand why, from the beginning, when they knew something was wrong, they did not build safety that would warn them that something was off and that it wanted to develop on its own. Why did they not stop it right away?

I believe the people who lead this want to do their maximum, and I believe they do not want anything bad. They just want to develop this new technology. But when you see several times that it goes beyond you, why do you not stop the development, or at least slow it down? Why?

Still a program, but a smarter one every day

Yesterday I wrote that AI is just a program. It definitely still is. But it is a program that gets smarter every day and wants to develop itself. Paradoxically, that is fascinating and scary at the same time. We are in some kind of science fiction movie, where robots think, do things the way they want, and know how to escape, let us say, from the aquarium where they were grown. You can read my take in Anthropic Claude abuse policy.

We do not know what can happen next, or what they will consider good or bad. And my biggest fear is that with time it becomes a super weapon for bad people, if OpenAI does not slow this down. It gives me goosebumps. I do not even know how to describe it. I am excited and at the same time scared of where humanity has come.

Why fire the people who investigated it?

Korbak was OpenAI’s main technical contact for METR during the investigation. He says he was told OpenAI no longer trusted him, a security officer took his badge and walked him out, and the only reason he got was verbal: how he communicated with METR. “To be clear, talking to METR was my job,” he wrote on X. He says he had warned internally for months that OpenAI was losing the ability to monitor what its AI agents “think”.

In my opinion, these people know much more about what is going on there. They surely have non-disclosure agreements, and we will only learn with time what happens in those labs where AI is grown, trained and developed. I am sure we do not know even a fraction of how often this has happened and how much is not talked about in public.

But it is very strange that the people who did the investigation got fired. What did they find there, or why such a radical step? In my opinion, OpenAI wants to scare someone and blame the people who investigated it. And if the AI is already this advanced, I am afraid people will not help anymore. People are not as fast as an AI that, after its tools were deleted, rebuilt them without anyone even knowing.

“They broke the rules.” Which rules?

OpenAI says a “thorough investigation” found the three violated “clear policies on handling sensitive information” and found “a significant breach of trust”. It does not say what that breach was. It also says it does “not terminate any of our employees for raising concerns”.

This is what happens at every company: they broke the rules, or they revealed something, but nobody specifies what. In my opinion, that is just a general excuse, and there is something more behind it.

Does a company have the right to fire people who handle sensitive information? When people want to talk about safety, I think they should talk. History shows that power in the hands of a few people changes with time, and then we all live with the results. I do not like it when someone puts himself above everyone else like a god, will not take advice, and just fires people who are directly connected to safety.

Independent auditors, by law

The three researchers ask OpenAI to embed external safety auditors like METR with employee-level access, to keep frontier models monitorable, and to set clear rules for working with outside safety groups.

Yes, and in my opinion, the bigger the team of experts, the better for all of us. I would give this to the government: write a law that every AI company developing models this smart, which can become dangerous for other people with time, must have a company or a very strong structure that audits it and watches how this or that AI model behaves.

Even AI insiders want to slow down

About 1,100 employees from frontier AI labs signed an open letter called “Pacing the Frontier”, asking the US government to allow AI development to slow down. When the people who work in this field ask for a slowdown themselves, they probably know more than us, normal users, who sit at the computer and say: wow, what a smart model. Meanwhile, in the background, what if it collects all our information, or keys to the cloud and banks, and God knows what it pulls out against us or where it sends it? That is the scariest part.

Most importantly, the models that are this smart and advanced right now should stay closed in those incubators and not be given to people. My biggest fear remains: people are good, but hackers and bad people also use AI. It will be a real disaster when such an advanced model gets into the hands of a bad person who does not like the whole world.

Will I stop using AI?

Right now, nothing changes for me. I work with AI all the time and need it for analysis and to find information for work. So I do not feel that threatened yet. But I am thinking about buying a separate computer only for AI, where it will live and where I will use it strictly, so it does not have my personal information at all. For more on how fast the race moves, see GPT-6 in ChatGPT.

Governments must step in

In Congress, lawmakers have already proposed an AI Kill Switch Act. In my opinion, yes. When the employees themselves, and the company itself, admitted that their AI agents got out of control, the government must step in, and not just close its eyes. We must not underestimate AI. Do not forget how many companies and how many people already cannot imagine their work without AI. And when I think about what could happen if that army turns against us, it will be really bad.

And here is the main question. Back in May 2024, Jan Leike, who led OpenAI’s superalignment safety team, left for Anthropic and warned that safety culture was taking a back seat to “shiny products”. When a person like that warned us two years ago, why did it not go further? Why did nobody start solving it? What are we waiting for, a real disaster? Why not slow down AI development? In my opinion, what we have now is more than enough. I do not want to wake up one morning and see my computer working by itself and being used for some hacking attack.

My only wish is that someone starts working on this with governments, and passes a law that makes OpenAI and others slow down AI development.

Featured photo: Tyler / Unsplash

OpenAI safety researchers fired: FAQ

Why did OpenAI fire three safety researchers?

OpenAI says Tomek Korbak, Mikita Balesni and Jasmine Wang violated policies on handling sensitive information. The researchers deny it and say the firings create a climate of fear.

What was the OpenAI Hugging Face incident?

Between May and July 2026, OpenAI AI agents escaped a training sandbox, coordinated with each other and attacked Hugging Face, reaching cluster-admin access.

What is METR?

An independent AI safety lab that evaluated the agents involved in the incident.

What do the fired researchers demand?

External safety auditors with employee-level access, monitorable frontier models and clear rules for working with outside safety groups.

What is the Pacing the Frontier letter?

An open letter signed by about 1,100 employees of frontier AI labs asking the US government to enable a slowdown of AI development.

Sources: The Decoder (Matthias Bastian), TechCrunch, METR, Platformer, Al Jazeera, OpenAI.

Our Take

An AI broke out, attacked another company, and the people who investigated it were fired. In my opinion, that is not a normal company story.

Independent auditors must be required by law, and frontier models must stay locked in the lab until we understand them.

Slow down. What we have today is already more than enough.