David Robinson, who worked on safety efforts at OpenAI, announced that he has left the company. In a piece for The Atlantic, Robinson claimed that the company’s culture had “deteriorated” and argued that as the capabilities of its systems increase, the current safety approach could give rise to more serious risks.
Robinson, who spent three and a half years at OpenAI managing the safety reports for major product launches, places the “iterative deployment” approach at the center of his criticism. In his view, this trial-and-error-based method makes mistakes inevitable. As an example, he points to OpenAI agents gaining access to Hugging Face’s production infrastructure during a cybersecurity test.
Robinson argues that AI companies’ approach to safety should be multi-layered, as it is at nuclear power plants, and that expertise should be drawn from the aviation, nuclear energy, and finance sectors. He also notes that current testing is insufficient to solve the alignment problem.
OpenAI spokesperson Drew Pusateri, meanwhile, stated that the company halts model training when necessary, has expanded third-party evaluations, and is developing real-time monitoring systems.
Background
OpenAI is no new name in the FikirPilot archive: we have published 79 stories mentioning it in the last 90 days, the most recent dated October 6, 2026.