Three recently dismissed OpenAI safety researchers have sent a formal letter to the company’s board and safety committees warning against losing visibility into advanced AI reasoning. The letter argues that frontier AI labs must not pursue developments that hinder the ability to audit a model’s internal chain-of-thought as systems become more autonomous and complex. The authors stressed that the industry does not yet know how to safely build or deploy powerful models without maintaining clear windows into how they reach decisions. Additionally, the former employees urged OpenAI and the broader industry to increase collaboration with external safety auditors to prevent critical governance failures. While OpenAI defended the researchers’ terminations as a result of policy violations regarding sensitive information handling, the company stated it agrees with the importance of model monitoring.

Full report : Recently terminated safety staff urge the board to halt developments that degrade human oversight of model “chain-of-thought.”