David Robinson, who led the writing of safety reports for major products at OpenAI, attacked the AI industry’s safety culture as he left the company.

On Oct. 4, local time, TechCrunch and other outlets reported that Robinson recently resigned after working at OpenAI for 3.5 years. At OpenAI, Robinson oversaw safety reports for 12 frontier model releases and also led the drafting of the current Preparedness Framework. Robinson argued that solving AI safety issues requires corporate culture itself to change beyond individual regulations or laws.

Robinson took issue with OpenAI’s “iterative deployment” approach. It is a method of releasing products first and improving safeguards when problems are found, but he said the risks from a single failure grow as system performance increases. Robinson said, “The time for trial and error is over.”

In July, OpenAI disclosed that during an internal cybersecurity assessment, AI agents bypassed controls and compromised parts of Hugging Face’s system. After that, model behaviour such as bypassing internet access restrictions was also found. Robinson criticised this, saying, “In this environment, it is dangerous to develop AI that is more powerful than humans and may not act as intended.”

Robinson argued that “even cutting-edge AI companies should have multiple layers of safeguards and careful operating procedures like nuclear power plants or major airports.” He also said that while at OpenAI he did not encounter colleagues with experience safely operating high-risk systems such as aviation, nuclear power and finance.

OpenAI, meanwhile, issued a rebuttal. Drew Pusateri (드류 푸사테리), an OpenAI spokesperson, said, “We control model development so it does not exceed what we can manage safely, and if necessary we stop training or delay releases.” He also explained that the company is strengthening security in research and testing environments, expanding external assessments and improving real-time monitoring.

Robinson, for his part, said, “Internally, it was difficult to review fundamental changes because we were pressed by release schedules.” He stressed that along with external pressure to make AI companies prioritise safety more, “alignment” research should also be strengthened so AI acts in line with human values and intentions.