Fired OpenAI safety researchers publish open letter on AI oversight

Three OpenAI safety researchers lost their jobs last week. Now they have published a letter criticizing the company for putting profit ahead of safety.

The researchers are Jasmine Wang, Tomek Korbak, and Mikita Balesni. OpenAI says it fired them for sharing confidential material outside the company. They say they were doing their jobs.

What happened and when

The terminations took place on or around October 1, 2026. OpenAI cited violations of company policies on sensitive information.

An internal investigation reportedly found that the trio had passed confidential materials to an outside AI safety organization. The researchers disputed that framing. They argued their conduct fit within their professional responsibilities.

That argument carries some weight given Korbak’s former role. He previously served as OpenAI’s liaison to METR and Redwood Research, two external safety groups. When part of your job description is talking to outside safety researchers, the line between “collaboration” and “leak” can get blurry fast.

On October 8, the three sent a letter to OpenAI’s board and its safety committees. It has since been published on Balesni’s personal website.

The letter’s central warning concerns something called “chain of thought” monitoring. Advanced AI models can produce a trail of intermediate reasoning before arriving at an output. Safety researchers watch that trail for signs a model is doing something it shouldn’t. The letter highlights potential risks if that monitoring capability is lost.

The researchers also pointed to broad uncertainty across the industry about how to safely build and deploy models that operate beyond human oversight. Their proposed fix was twofold. They called for industry-wide collaboration on model monitoring and for independent safety auditors to be brought in.

OpenAI’s response

OpenAI has drawn a firm line between the letter and the firings. The company said the terminations were not connected to the researchers’ safety advocacy.

It went further. OpenAI said it agrees with the letter’s recommendations. In its view, the dismissals were necessary because of breaches of trust, full stop.

A pattern of safety exits

The firings did not happen in a vacuum. Just days earlier, David Robinson resigned from his post as OpenAI’s safety transparency lead.

Robinson had been critical of the company’s internal culture. He raised concerns about how OpenAI balances rapid technical progress against adequate safety measures.

There is also history here. In 2024, OpenAI fired researchers over similar violations, also amid ongoing safety concerns. The company has additionally faced public criticism of its safety practices, including recent security incidents.

What this means

For the wider AI industry, the letter’s substance matters more than the personnel drama. Chain of thought monitoring is one of the few windows researchers currently have into how advanced models reach their conclusions. The call for shared monitoring standards suggests the authors see it as a collective problem that no single company can solve alone.

There is a structural tension worth watching too. Labs want outside validation of their safety work, which requires sharing information. They also want to protect proprietary research, which requires keeping it close. The firing of a former liaison to external safety groups for allegedly sharing material externally sits right on that fault line.

The next signals to watch are fairly concrete. Does OpenAI announce any formal arrangement with independent auditors? Do other labs respond publicly to the letter’s call for joint work on model monitoring? And do more safety staff follow Robinson and the three researchers out the door?