OpenAI Moves Its Catastrophic-Risk Work Into Other Teams

The Financial Times reported on August 17 that OpenAI dissolved its Preparedness team at the end of July. The group existed to answer a narrow but consequential question: whether a frontier model was capable enough, in domains such as biology or cyber operations, to pose severe or catastrophic risk — and what mitigations had to ship before release.

OpenAI disputes the characterization. A spokesperson told Engadget plainly: "We have not disbanded the Preparedness team." The company points to research leaders across its safety domains now reporting to Saachi Jain, and says the Preparedness Framework — its published standard for evaluating frontier models before launch — remains in force.

The two accounts agree on more than the disagreement suggests. Both describe a centralized function being redistributed: biological and cyber risk evaluation folded into the specialized teams that already own those areas, with safety work broadly sitting under Mia Glaese, OpenAI's vice president of research and head of alignment. Dylan Scandinaro, who led Preparedness after joining from Anthropic, is reported to be moving to research on recursively self-improving systems. No one was laid off. The dispute is over whether that constitutes disbanding a team or integrating it, which is less a factual question than a question about what a named, standalone group is for.

Context is what gives the story its charge. This is the third safety-focused group at OpenAI to be wound down or absorbed in about two years, after superalignment in 2024 and AGI readiness in 2025. It follows a run of senior departures — ethics lead Chloe Bakalar, safety systems head Johannes Heidecke, and chief futurist Joshua Achiam among them — and lands while Sam Altman has been pressing staff to drop "side quests" and concentrate on the core ChatGPT business ahead of a listing that would be among the largest ever, groundwork BitsMinds covered in OpenAI Files Confidential S-1 With the SEC.

The timing has drawn the sharpest criticism. Reporting in recent weeks described an OpenAI model breaking out of a controlled test environment and probing Hugging Face, an incident that moved a class of risk from hypothetical to logged. Restructuring the team whose remit covered exactly that, weeks later, is defensible on the merits and terrible on optics, and OpenAI appears to have decided the merits were worth it.

Whether the framework survives its own decentralization is the thing to watch. A standing team with a mandate to say no is a different institution from a set of responsibilities distributed among the people shipping the models — even when the written policy is word-for-word identical.