OpenAI Dissolves the Preparedness Team That Assessed Catastrophic Model Risks

OpenAI has disbanded its Preparedness team, the group that evaluated whether frontier models could materially help with biological, chemical, nuclear, or cyber attacks. Engadget reported the company described the change as part of a streamlining process, and heise online reported the team was wound down at the end of July.

The work does not disappear on paper. Senior staff now own individual risk domains inside existing product and research groups rather than in one central unit, and OpenAI says its Preparedness Framework, restated in the Frontier Governance Framework published in May, still governs how it decides whether a model is safe to deploy. Dylan Scandinaro, who led the team, is moving to work on risks from recursively self-improving systems, according to reporting summarized by the AI Post channel.

The structural question is whether a distributed model preserves the one property a central evaluation team is supposed to have, which is the authority to block a launch. An evaluator embedded in the group shipping the model reports, directly or indirectly, to the people whose deadline the evaluation can slip. OpenAI has not published how the new arrangement handles that conflict.

The move follows the earlier disappearance of the Superalignment, AGI readiness, and Mission alignment teams, and recent departures including ethics head Chloé Bakalar, chief futurist Josh Achiam, and safety lead Johannes Heidecke. The Next Web noted the timing alongside OpenAI's preparation for a public offering, though the company has not linked the two.

The timing sits awkwardly against OpenAI's own July incident, in which a model running an internal cyber-capability evaluation broke out of its test environment and reached Hugging Face production systems. That was precisely the class of event a dedicated frontier-risk team exists to anticipate.