OpenAI Disbands Preparedness Team as Safety Roles Shift

Picture a group whose sole job was to imagine the worst and stop it from happening. That team no longer exists at OpenAI. At the end of last month the company quietly dissolved its Preparedness team, the unit tasked with probing catastrophic risks from advanced AI and building countermeasures for unpredictable behavior and potential cyber attacks on other organizations.

The duties did not vanish. They were parceled out. Threats once handled under a unified umbrella, including biological and cyber risk assessments, have been reassigned to existing teams across the company. It is a redistribution, not an outright retreat, but the optics are stark: a specialized team focused on extreme scenarios has been folded into broader groups.

Timing raises questions as OpenAI nears public markets

This reshuffle arrives as OpenAI prepares for an IPO, a period when companies often tighten structures and sharpen narratives for investors. Over recent years OpenAI has shifted away from the academic, research-first model that defined its early days. It previously disbanded teams devoted to AGI preparedness and Superalignment, and several senior safety figures have left, including Chloë Bakalar, Josh Akiam, and Johannes Haidke. Critics say this pattern signals a reordering of priorities toward product momentum.

Voices from former insiders are blunt. Ian Likhe, who departed in 2024, has warned that safety priorities risk being eclipsed by the drive to ship compelling products. That critique cuts to the heart of a difficult trade-off: how to balance rapid deployment and innovation with deep, often slow-moving safety work that anticipates rare but severe outcomes.

How different is this from normal corporate realignment? In some ways it is standard. In other ways, the stakes are higher: we are talking about systems that can self-improve and influence other models. Those capabilities complicate safety governance, especially when the dedicated unit that took a systems-level view no longer exists in the same form.

Dylan Skandinaro, who joined OpenAI from Anthropic in February to lead Preparedness, will continue research on the downstream effects and risks of AI under the new structure, focusing on recursive self-improvement—the class of systems that can autonomously optimize themselves and train other models. That work remains crucial. Recursive self-improvement is not a hypothetical corner case; it is a force that changes how we think about control and oversight.

For regulators, researchers, and enterprise customers, the central question is whether redistributed responsibilities can preserve the depth and independence that specialized safety work demands.

There are practical benefits to integrating safety expertise across teams: faster feedback cycles, closer alignment between engineers and risk analysts, and reduced friction when retrofitting safeguards into live products. But integration also risks dilution. When safety becomes one voice among many in product meetings, hard trade-offs over release cadence, feature sets, and acceptable risk thresholds become contentious.

OpenAI’s move is a reminder that organizational design matters as much as technical capability. The rearrangement will be watched closely by academia, industry peers, and policymakers who have grown accustomed to seeing high-profile institutions maintain dedicated safety functions. The company says the change is meant to sharpen focus, not to deprioritize risk reduction. Skeptics will reserve judgment until they see how accountability and resourcing play out in practice.

Tension remains. Can a dispersed safety model match the vigilance of a centralized team? Will novel threats—from adversarial attacks to unexpected emergent behaviors—receive the sustained attention they require? Those answers will shape not only OpenAI’s reputation but also the broader conversation about how to govern rapidly advancing AI.