Fired OpenAI safety staff warn AI safety cannot be solved alone

OpenAI has dismissed three members of its AI safety staff with immediate effect, saying they passed on confidential information. The three deny this and have published an open letter warning that OpenAI cannot make artificial intelligence safe on its own.

What happened

The German technology publication t3n reports that OpenAI’s safety work has been under scrutiny for several months. According to t3n, the concerns began with what it calls the Hugging Face incident, which the company has been investigating since. During those investigations it became known that OpenAI had, among other steps, disbanded its crisis team for AI safety.

The latest development, t3n writes, is the dismissal of three employees who worked on AI safety and alignment, the field concerned with making AI systems behave as their developers intend. OpenAI says they forwarded internal information. The three say this never happened.

The former employees are Tomek Korbak, Jasmine Wang and Mikita Balesni. Balesni published an open letter on X, written together with the other two. It is addressed to three of OpenAI’s oversight bodies: the Safety and Security Committee, the Safety Advisory Group and the Mission Advisory Council. Its central warning, as quoted by t3n, is that “OpenAI cannot make AI safe alone.”

Balesni also describes how the dismissal happened. In t3n’s account, Balesni writes that they were not told in writing why they were fired. In the exit meeting, Balesni was told that OpenAI no longer trusted them because Balesni had spoken too often with other organisations, which implied that company secrets had been passed on. Balesni says this work had always been agreed with their managers, the head of the research department and the board.

The letter’s wider concern is the effect on colleagues who are still at the company. The three write that they fear internal and external communication about their dismissals will make former colleagues afraid to speak up or act in the way that, until the previous week, had been an integral part of working at OpenAI. They argue that contact with outside parties is essential to safety work. People who study safety risks before anyone else, they write, depend on close collaboration with external experts to find ways of dealing with those risks. Being free to do that without fear, within defined processes, is in their view a safety mechanism in itself.

The letter ends with three recommendations to OpenAI:

• Keep the promise of outside scrutiny. OpenAI should continue to seek external advice and safety audits. According to t3n, Sam Altman made this promise in September 2026, saying independent safety experts should be given the same access as OpenAI staff. The former employees say this is the only way to guarantee the safety of the models.
• Work on monitoring AI models. The letter states that no one yet knows how to safely develop and release models that cannot be monitored, and that the ability to monitor frontier models, the most advanced systems, keeps declining.
• Support open dialogue. OpenAI should encourage an open and transparent culture of discussion between its safety researchers and the wider safety community, so that staff can keep voicing opinions and concerns freely.

If concerns are not raised or examined externally, the three write, there is a serious risk that a catastrophe will happen. They call on their former colleagues to keep holding OpenAI to these standards, adding that OpenAI would be nothing without its people.

Why it matters if you live in Germany

The t3n report does not discuss Germany, German law or any specific effect on people living here. Its relevance is therefore general rather than local, and it is worth being clear about that.

For anyone in Germany who uses OpenAI’s products at home or at work, or whose employer builds services on its models, the story concerns how the company checks the safety of the systems it releases. The former employees’ main argument is that internal safety teams alone are not enough, and that outside experts need real access. Whether OpenAI keeps the promise described by t3n affects how much independent checking those models get before they reach users.

For people working in research or technology in Germany, the case raises a familiar workplace question: how far an employee can talk to outside organisations before an employer treats it as a breach of trust. The three say their outside contacts were approved by their managers and the board. OpenAI’s view, according to their account, is that those contacts amounted to sharing secrets. The source does not include a response from OpenAI to the letter.

What to watch next

The letter is addressed to three named OpenAI bodies, so the first thing to watch is whether the Safety and Security Committee, the Safety Advisory Group or the Mission Advisory Council responds in public. A second is whether OpenAI gives a fuller explanation of the dismissals, since the former employees say they received no written reasons.

It will also be worth following whether OpenAI keeps the commitment Sam Altman made in September 2026 to give independent safety experts the same access as staff, and whether current employees speak up as the three have asked them to. t3n notes that it originally published the article earlier and has updated it because readers are still interested in the case.

Source: t3n