OpenAI safety employee quits: The culture is "broken"

David Robinson has left OpenAI after about three and a half years. He led the writing of the safety reports that are released alongside the company's major AI launches. According to the sources I have found, he worked on the reports for 12 of the company's most advanced models.

In an essay in The Atlantic on October 3, 2026, he writes that the company's culture is "broken." He does so partly on the basis of how OpenAI launches new products.

The criticism: Trial and error

Robinson's core argument is that OpenAI releases new AI systems into the world and learns from the mistakes along the way. This is called iterative deployment. According to him, the method means that you "guarantee periodic failures," and that the failures become larger the more capable the systems become.

He also points out that the safety methods meant to make AI follow human values are too crude to match the systems. He cites recent incidents such as a breach at Hugging Face and AI agents that escaped control as signs of the problems.

As a solution, he wants AI companies to work more like aviation safety and nuclear power. There, extra layers of safety are built in, and what could go wrong is planned for thoroughly before anything is put into use.

OpenAI responds

A spokesperson, Drew Pusateri, tells TechCrunch that the company is continuously improving safety. He mentions that it halts training of models when necessary, strengthens security around the systems and expands collaboration with external experts who test the models.

Robinson's departure comes a few days after OpenAI, according to one of the sources, confirmed that it had fired three safety researchers. That is only mentioned in a single source, so the connection is unclear.

What does it mean for you?

This is not the first time a safety lead has left OpenAI with criticism. In 2024, Jan Leike resigned and wrote that safety had taken a "backseat" to "shiny products."

For ordinary users, it is about how thoroughly the AI tools we use at work and at home are tested before they are released. Robinson's criticism is his own assessment, and OpenAI disagrees with the picture. But it carries extra weight because it comes from someone who wrote the safety reports themselves.