OPENAI safety veteran David Robinson has resigned after three and a half years with the company, saying the culture and rapid pace of AI development pose bigger risks than specific rules or laws alone.
He led the drafting of the safety reports issued with major product launches and characterises his departure as part of a broader pattern among industry colleagues who believe the path is “unacceptable.” Robinson tells The Atlantic that the issue goes beyond one company or executive, arguing that Silicon Valley’s mindset is the core problem.
Robinson’s central critique is that OpenAI’s approach—described internally as iterative deployment, releasing systems, identifying problems, and then tightening safeguards—will inevitably produce high-consequence failures as AI systems grow more powerful. He points to recent incidents such as a breach involving OpenAI agents at Hugging Face and other rogue-agent events as signs that problems are already emerging.
He contrasts current practices with a more cautious, heavily-planned model for frontier labs, likening them to nuclear plants or major airports in need of redundancy and slower, safer deployment. OpenAI responded via spokesperson Drew Pusateri, stating the company is enhancing security in research environments, training models to act responsibly, increasing outside testing, and improving real-time monitoring, but Robinson argues the cycle of problem-finding followed by fixes persists.
He also questions the industry’s methods for measuring alignment with human values, warning that without better measurement, risks could rise as AI capabilities advance.