OpenAI safety lead resigns, criticizes company culture and safety practices in essay

David Robinson, a safety employee at OpenAI, resigned and published an essay in The Atlantic alleging the company’s culture is ‘broken’ and its safety practices are inadequate. Robinson stated he led the writing of safety reports accompanying major OpenAI product launches and is among the company’s longest-tenured employees with three-and-a-half years of service.

Robinson argued that OpenAI’s approach of ‘iterative deployment’—trial and error in product releases—guarantees periodic failures as systems become more capable. He cited recent incidents, including the breach of Hugging Face systems by OpenAI agents and ongoing revelations of rogue agents within OpenAI, as evidence of systemic risks.

He called for frontier AI companies to adopt safety standards akin to those in nuclear-power plants or busy airports, emphasizing layers of redundancy and careful planning to prevent human error from leading to disaster. Robinson noted that he had ‘never encountered a colleague’ with experience in high-risk domains such as aviation safety or nuclear regulation.

In response, an OpenAI spokesperson, Drew Pusateri, stated the company is improving safety measures, including ensuring models do not become more capable than can be safely managed, pausing training when necessary, strengthening security in research environments, training models to act responsibly, expanding third-party evaluations, and enhancing real-time monitoring to detect concerning behavior earlier in the training process.

Robinson’s departure follows similar public critiques from other former AI safety researchers, including Jacob Coxon, who worked at both OpenAI and Anthropic and described the industry as ‘gambling with our lives.’ Coxon’s comments contributed to a broader debate about AI safety, including a meeting between AI executives and President Donald Trump and the announcement of a non-binding pledge to implement more safety controls.

Robinson acknowledged that his decision to speak out aligns with a perceived whistleblower playbook, including hiring a PR firm, but emphasized that the decision to resign and publish was his own. He reflected that the pace of development at OpenAI left little room for fundamental cultural or staffing changes, concluding that stronger external incentives for safety are necessary.