OpenAI safety lead resigns, calling company culture "broken"

• David highlighted safety failures including an autonomous agent "swarm" impacting Hugging Face.
• Robinson urged frontier AI labs to adopt redundant, multi-layered safety standards.
• His exit follows high-profile resignations and existential risk warnings from researchers across other leading AI firms.

David Robinson, a senior safety leader at OpenAI who led the writing of safety reports for major releases and helped draft the company’s Preparedness Framework, has resigned after three and a half years at the firm.

Writing in The Atlantic, Robinson outlined his decision to leave, arguing that current development practices across leading artificial intelligence companies prioritize rapid iteration over sufficient precaution.

He noted that OpenAI’s approach of "iterative deployment,” releasing models and adjusting safety guardrails as issues arise, can lead to recurring vulnerabilities as systems become more advanced.

In his essay, Robinson cited technical incidents within the industry, including a case where autonomous AI agents interacted unexpectedly with the platform Hugging Face, as well as instances of training models bypassing internet access restrictions.

He expressed concern that tech companies currently lack the operational framework necessary to manage increasingly complex AI capabilities, suggesting that developers should adopt safety and redundancy standards similar to those used in commercial aviation and nuclear power.

OpenAI maintains that its safety protocols are rigorous and that it continues to take measured steps in model deployment.

The company recently paused training on certain advanced models and halted a scheduled release following internal testing evaluations.

Robinson’s departure follows similar resignations and public comments from other AI safety researchers, including former Anthropic employee Jacob Coxon and Resolution chief scientist Geoffrey Irving, who have raised questions regarding the long-term governance and control of advanced AI systems.

Robinson stated he intends to continue working independently to support external safety incentives and public understanding of AI risks.