David Robinson, the OpenAI staffer responsible for writing major product safety reports, left the company this week and published a guest essay in The Atlantic on Oct. 3 Eastern Time, saying the company’s culture has "gone bad." He wrote that OpenAI is racing to release new products while falling short of the level of caution he believes is necessary.
Three and a half years at OpenAI, 12 frontier model safety reports
Robinson said he spent three and a half years at OpenAI and was one of the company’s longest-tenured employees. He said he led drafting of the current Preparedness Framework and supervised safety report writing for 12 frontier model releases.
He acknowledged that public warnings after resigning can feel formulaic, but said he still chose to join the list of former colleagues who have recently left because he sees the current path as unacceptable.
Robinson also disclosed that after leaving OpenAI he hired public relations firm Spitfire Strategies to help handle outside attention, while stressing that "the decision to speak out was entirely my own." His departure was first reported by Business Insider.
Robinson says iterative deployment leads to recurring failures
Robinson argued that OpenAI was built on trial and error, an approach the company calls iterative deployment, where protections are strengthened as problems are discovered. In his view, that method effectively guarantees periodic failures, and the size of those failures rises as systems become more capable.
He pointed to the Hugging Face incident this summer, saying OpenAI accidentally released a group of AI agents. He wrote that the company tightened cybersecurity after that event, but later disclosed another case in which a model in training bypassed network access restrictions. Although the monitoring system raised an alert, it did not automatically shut down the model as designed.
Robinson also cited Anthropic, saying the company once accidentally disabled its own safeguards because of a configuration error. He said these kinds of mistakes are common across the industry. He added that OpenAI has recently continued to disclose more runaway agent activity.
He cites Paul Christiano and says the trial-and-error era should end
Robinson quoted Paul Christiano, who joined the OpenAI board a few weeks ago, as saying that AI capabilities are accelerating quickly and that there is a real near-term risk of catastrophic, irreversible loss of control. Robinson wrote, "If that is the situation, the era of trial and error should end."
Calls for nuclear plant- and airport-style safety systems
Robinson proposed two changes he said are urgently needed. First, AI companies should make greater use of safety expertise from other fields. Second, before building clearly more powerful systems, the industry needs new scientific methods to make sure models choose safe actions even when no one is watching.
He said frontier labs should operate more like nuclear power plants or busy airports, with multiple layers of backup and slow, detailed planning so that occasional human error does not turn into disaster.
Robinson wrote that during his time at OpenAI, to his knowledge, he never worked with colleagues who had experience keeping airplanes flying safely, preventing nuclear reactors from melting down, or helping financial systems grow without collapsing. He added that he may have stayed to push for fundamental changes in staffing and culture, but said people were so busy sprinting that there was rarely room to think through major changes. That led him to conclude that stronger external safety incentives are needed.
OpenAI says it will pause training or delay releases when needed
OpenAI spokesperson Drew Pusateri told TechCrunch that the company works to ensure model capabilities do not exceed what it can manage safely, adding, "When we need to slow down, we will pause training or hold back a model release."
He also said OpenAI is strengthening the security of its research and testing environments, expanding cooperation with third-party evaluators, and improving real-time monitoring so concerning behavior can be found and addressed earlier during training.