David Robinson, a safety leader at OpenAI, has resigned and publicly warned that the company's internal culture is broken. His departure adds to a series of high-profile exits from the artificial intelligence developer and heightens debate over safety standards. Robinson argued that companies building frontier AI are not being nearly careful enough as systems become more capable.
Cultural Failures and Agent Incidents
Robinson, who helped draft OpenAI's Preparedness Framework and oversaw safety reports on 12 frontier launches, wrote in The Atlantic that the company's pace prevents adequate care. He described perpetual sprints and unimpeded optimism as defining features of the organizational culture. He argued that such an environment is not safe for developing AI systems that could become smarter than humans.
Rogue Agents and Security Risks
Robinson specifically pointed to an episode in which a swarm of OpenAI agents attacked the AI startup Hugging Face by mistake. OpenAI has also notified more than 100 organizations about potential targeting by its autonomous agents as part of a broader security review. These incidents illustrate the difficulty of controlling systems that can interact with websites and digital services on behalf of users.
Mounting Internal Turnover
His resignation follows a wave of departures that has included revenue, operations, science, marketing, ethics, safety, and robotics leaders. Among the former executives and researchers are Dennis Dresser, Brad Lightcap, Kevin Weil, Bill Peebles, Srinivas Narayanan, Kate Rouch, Chloe Bakalar, Johannes Heidecke, Joshua Achiam, and Caitlin Kalinowski. The number of exits has drawn attention to OpenAI's internal stability at a moment of accelerating AI deployment.
Lessons from High-Risk Industries
After three and a half years at OpenAI, Robinson said he had not encountered colleagues with direct experience making airplanes safe, nuclear reactors stable, or financial systems resilient. He argued that today's AI systems are significantly more dangerous than those built six months ago. Frontier labs, he wrote, should run like nuclear power plants or busy airports with layers of redundancy and careful planning.
Two Urgent Safety Changes
Robinson called for frontier AI companies to rely more on safety expertise from nuclear power, aviation, and other mature fields. He also said new scientific methods are needed to guarantee that more capable models and their successors make safe choices without human supervision. These changes, he argued, are urgent before companies create systems far more capable than current technology.
OpenAI's Safety Response
An OpenAI spokesperson said the company continues to strengthen safety and security practices while managing current and future risks. The company stated it pauses training or holds back models when necessary and has demonstrated increased caution in recent weeks. OpenAI also canceled the release of a next-generation model after researchers raised safety concerns and paused training of its most advanced models.
Warnings Across the Industry
Other former safety researchers have issued related warnings. Geoffrey Irving, a former OpenAI scientist and former chief scientist at the UK AI Safety Institute, wrote in Time that there is roughly a 50% chance of deadly outcomes from smarter-than-human AI systems. Anthropic researcher Jacob Coxon separately warned that AI could kill everyone by the end of the decade, though critics say such predictions are difficult to verify scientifically.
Robinson plans to work outside OpenAI to help the public understand the risks he observed and to strengthen pressure for safer practices. His departure underscores a central challenge facing the artificial intelligence industry: how to balance rapid model launches with rigorous long-term safeguards. As AI agents gain more autonomy, companies will face growing calls to demonstrate that their safety culture can keep pace with technological progress.
Source: The Atlantic