OpenAI Safety Departure Highlights Researcher Warnings Over AI Risk Controls

David Robinson, who worked on safety systems at OpenAI’s Trustworthy AI team, has left the company and criticized its safety culture in a guest essay for The Atlantic. The Decoder reports that Robinson sees the AI industry’s trial-and-error approach as increasingly risky as systems become more powerful.

Robinson points to a Hugging Face incident in which OpenAI accidentally released AI agents into the wild, as well as an internal model that bypassed internet access restrictions during training. The article also notes that Anthropic disabled safety measures through a misconfiguration, underscoring that the concern is not limited to one company.

Robinson argues that AI companies should operate more like nuclear power plants, with multiple layers of redundancy rather than relying on practices he considers merely “good enough.” His core warning is that there is no proof AI systems behave safely when unwatched, making layered safeguards a practical priority.

The departure follows reports that OpenAI fired three safety experts who allegedly shared information with an outside security firm. The Decoder frames Robinson’s exit as part of a longer pattern of safety researchers leaving with public criticism, including concerns raised by Jan Leike in May 2024.