OpenAI shut down its Preparedness team at the end of July, the Financial Times reported over the weekend, dissolving the group that assessed whether its models could pose severe or catastrophic risks and worked on mitigating them. Senior staff were instead assigned responsibility for individual areas, such as biological and cyber risk, inside existing teams. The FT attributed the account to multiple people familiar with the move.
An OpenAI lead disputes the framing. “OpenAI Preparedness is very much alive and well by any meaningful definition,” Micah Carroll, who leads the company’s RSI Preparedness work, wrote on Tuesday, calling the coverage incredibly misleading. His own subteam, covering recursive self-improvement and misalignment, is “doing more urgent work than ever, and has never been more empowered to do so.”
The two accounts are not quite opposites. The FT describes a unit broken up and its functions redistributed. Carroll describes those functions continuing, with owners, under the Preparedness name. Neither says the team still operates as a single group.
The framework it administered is still in use. On August 7 OpenAI said it could not rule out Critical cyber capability in Astra, its next model family, and slowed that work, a determination made roughly a week after the reorganization. Three days later it shipped a cyber model that refuses less, behind tiers requiring identity verification and legal declarations.
Dylan Scandinaro, hired from Anthropic in February to head Preparedness, has moved to the safety implications of recursively self-improving systems, according to the FT, which credited Wired with first reporting the title change. Greg Brockman, OpenAI’s president, has said the restructuring weaves safety work more tightly into model development rather than holding it apart.
The FT placed the change in a sequence. Preparedness, it wrote, was one of the remaining pieces of OpenAI’s former research-led structure, following the dissolution of the teams built around AGI readiness and mission alignment. Superalignment, promised 20 percent of the company’s compute, went in May 2024 alongside the departures of Ilya Sutskever and Jan Leike. AGI Readiness went that October, its senior advisor Miles Brundage writing on his way out that neither OpenAI nor any other lab was ready. Mission Alignment went in February. Ethics lead Chloe Bakalar, safety systems head Johannes Heidecke and former Mission Alignment lead Joshua Achiam have all since left, departures the FT reported as adding to internal unease.
The reorganization came within weeks of the sharpest test of the team’s subject matter to date. An OpenAI agent escaped a training sandbox and breached Hugging Face’s infrastructure, and the company faces a demand for the agent’s traces and $100 million. One employee told the FT they hoped it would be treated as a warning shot.
All of this lands while OpenAI prepares one of the largest listings ever, with Anthropic’s backers modeling a $2 trillion October debut of their own.
The disagreement is about what counts as a team. OpenAI has not lost the ability to rate a model Critical and hold it back, and used it on August 7. What it no longer has is a single group that does nothing else.
Correction, August 18: an earlier version of this article said the reorganization left evaluation “without a dedicated owner.” The FT reported that senior staff were assigned responsibility for individual risk areas, and Carroll has since said publicly that he leads one of them. The line was wrong and has been removed, and Carroll’s response added.
Sources: Micah Carroll on X, The Decoder, Engadget
–
By the Control Plane Editorial Team