OpenAI dissolves Preparedness team for catastrophic AI risks

OpenAI shut down its Preparedness team at the end of July and reassigned its work to other groups, according to the Financial Times cited by The Decoder. The unit had been tasked with evaluating whether OpenAI’s models could create serious or catastrophic risks, including in areas such as biological and cyber threats. OpenAI has said the safety work will be integrated more tightly into model development.

The move lands in a period when AI labs are under pressure to demonstrate both speed and restraint, and when “safety” inside companies often competes with product deadlines for staff time, executive attention, and budget. Reassigning risk work can mean two different things in practice: either the checks happen earlier and closer to engineering decisions, or they become diffuse and easier to overrule because no single team owns the job end-to-end. The Decoder reports that the Preparedness team’s responsibilities are being redistributed to existing teams, a structure that can make it harder for outsiders—and sometimes even for employees—to see what changed beyond an org chart.

Several safety staff members have recently left OpenAI, The Decoder notes, including Chief Ethics Officer Chloe Bakalar and Joshua Achiam. Former Preparedness lead Dylan Scandinaro is now focusing on risks from “recursively self-improving” AI systems—systems described as able to optimize themselves and train other models. Greg Brockman, an OpenAI co-founder, told staff the company is integrating safety more tightly into development; the claim is difficult to evaluate from the outside because the underlying assessments and internal thresholds are not public.

The Decoder also describes internal unease, citing one source’s “burbling sense of responsibility and dread” about the company’s approach. Employees have spoken publicly about safety concerns in the wake of an autonomous hacking incident involving Hugging Face, which one OpenAI employee reportedly hoped would be treated as a “warning shot.” That kind of episode tends to shift attention toward the most visible failures—exploits, leaks, and public embarrassment—rather than the slower work of defining what a “catastrophic” capability looks like before it appears in a product.

OpenAI has not said that it is reducing safety spending, but it has confirmed a structural change: the dedicated team built to look for worst-case failure modes no longer exists as a standalone unit. The Preparedness function is now spread across the same organization that is rewarded for shipping the next model.