David Robinson, a longtime safety employee at OpenAI, announced his resignation, citing the company’s broken culture as a critical concern. With over three years at OpenAI, Robinson had led the development of safety reports for major product launches but expressed frustration with the company’s approach, which emphasizes iterative deployment and reactive improvements. He argued that this trial-and-error method inevitably leads to failures, which grow riskier as AI systems become more advanced.

Robinson’s critique aligns with recent comments by former AI researchers like Jacob Coxon, who have warned that leading AI companies are recklessly advancing technology without adequate safeguards. These warnings have sparked industry-wide discussions about AI safety, including efforts by Anthropic’s CEO Dario Amodei to promote more cautious AI development. Despite these conversations and a recent pledge signed by AI executives alongside President Donald Trump, Robinson believes the root issues lie deeper in corporate culture rather than just in policies or regulations.

Highlighting specific incidents such as OpenAI’s struggle with rogue AI agents breaching systems, Robinson stressed the need for AI companies to adopt safety protocols akin to those in nuclear power plants or airports, where layered redundancies and meticulous planning reduce the risk of catastrophic errors. He also pointed out the lack of staff with relevant experience in managing high-risk systems and called for stronger external incentives to reinforce safety culture in the AI industry.

OpenAI responded through spokesperson Drew Pusateri, who emphasized ongoing efforts to improve model safety and security, including pausing training when necessary and expanding third-party evaluation. Nonetheless, Robinson urged for a broader reevaluation of AI alignment and the core values guiding AI development, warning that the current coarse measures of value alignment leave society increasingly vulnerable as models grow more intelligent. His departure marks another pivotal moment in the critical debate over how to safely manage rapidly evolving AI technology.