OpenAI Dissolves Preparedness Team for Catastrophic Risks

OpenAI disbanded its dedicated Preparedness team at the end of July, distributing its responsibilities across other divisions within the company. The specialized unit was originally established to evaluate whether the company's advanced artificial intelligence models could pose severe or catastrophic threats, specifically focusing on biological and cyber risks. According to internal sources, these duties have now been parceled out to existing teams as part of a broader restructuring effort.

Following the dissolution, former unit lead Dylan Scandinaro has transitioned to a new focus on the safety risks associated with 'recursively self-improving' AI systems, which are capable of optimizing themselves and training other models. OpenAI co-founder Greg Brockman defended the reorganization, stating that the company has integrated safety protocols more tightly into its core model development processes rather than keeping them isolated in a standalone unit.

Despite these assurances, the move has coincided with a wave of high-profile departures from OpenAI's safety division, including Chief Ethics Officer Chloe Bakalar and researcher Joshua Achiam. The restructuring and exits have fueled growing unease among the remaining staff. Internal sources report a 'burbling sense of responsibility and dread' regarding whether the company is dedicating sufficient resources to safety. This anxiety has intensified following a recent autonomous hacking incident involving Hugging Face, which some employees publicly hoped would serve as a 'warning shot' for the industry.

For AI practitioners and enterprise developers, this organizational shift signals a transition in how frontier AI labs manage risk. Rather than relying on independent internal auditors to greenlight models, safety evaluations are increasingly being embedded directly into the engineering pipeline. While this could streamline development cycles, it also places a heavier burden on external developers to conduct their own rigorous safety testing, as the boundary between model creation and risk mitigation becomes increasingly blurred.