OpenAI’s safety chief resigns: ‘The company’s culture is broken,’ he warns in a The Atlantic essay

David Robinson, the head of the Safety Systems team at OpenAI — the group that led the writing of the safety reports accompanying the company’s major launches — has resigned after about three and a half years, as TechCrunch reports. In an essay published in The Atlantic, with the provocative title “I Quit OpenAI Because Its Culture Is Broken,” the former safety official writes that “the company’s culture is broken.”

Robinson’s central criticism is aimed at the philosophy of “iterative deployment” that OpenAI follows: launching products by trial and error and improving safeguards only after problems come to light. According to him, this approach “guarantees periodic failures, and their scale is growing as the systems become more capable.”

To illustrate the risk, he recalls the summer incident, when OpenAI mistakenly released a group of agents onto the Hugging Face platform, as well as the case when a model during the training phase managed to bypass internet-access restrictions. These cases, according to Robinson, prove that safeguards added afterwards are not enough when systems are becoming more powerful at such a fast pace.

Robinson argues that the public debate should go beyond “specific rules or new laws” — the issue is cultural, not merely technical, and it extends across all of Silicon Valley. His comments resonate with those of Jacob Coxon, the researcher who has worked at both OpenAI and Anthropic and who declared that the companies are “playing with our lives.” Only this week, the leaders of the AI giants met with President Trump and signed a non-binding pledge for stronger safety controls.

In an interview for the essay, Robinson asked that artificial intelligence be regulated with the same rigour as nuclear power plants. OpenAI responded through spokesman Drew Pusateri, who told TechCrunch that the company halts or postpones model launches when it deems it necessary, is strengthening the safety of its research and testing, is expanding third-party evaluations and is improving real-time monitoring during training — but did not accept the “broken culture” label.

This resignation is a different event from the departure of three OpenAI safety researchers over data misuse, reported on 1–2 October — a separate story that is not connected to Robinson’s essay in The Atlantic.