OpenAI Disbands Preparedness Team Ahead of IPO

OpenAI disbanded its preparedness team at the end of July 2026, according to reporting by the Financial Times. The team's mandate — assessing whether models posed serious risks and developing mitigations — has been fragmented, with responsibility divided into domain-specific areas such as bio and cyber and folded into pre-existing teams. For developers building on OpenAI's APIs, this is a concrete governance change: the organizational layer that evaluated frontier model risk as a unified function no longer exists.

What the Preparedness Team Actually Did

The team existed to answer a specific class of question before model releases: could a given model cause serious harm at its capability level, and if so, how should that risk be bounded? That function sat between research and deployment, making it structurally distinct from product safety teams that handle post-release policy enforcement. Disbanding it and distributing its work into domain silos changes the incentive structure around those evaluations. Domain teams embedded in product or research orgs face different pressures than a team whose sole charter is adversarial risk assessment. Developers tracking OpenAI's trajectory as a deployment company should register this as a shift in how pre-release accountability is organized.

A Pattern of Safety Team Reductions

This is not an isolated restructuring. OpenAI has progressively dismantled its research-led safety infrastructure over several years:

Jan Leike, who resigned from OpenAI in 2024, told the Financial Times that the company was prioritizing "shiny products" over safety. That criticism carries specific organizational weight: Leike was a senior figure in OpenAI's safety research, and his framing maps directly onto the structural pattern the table above documents.

Dylan Scandinaro's New Focus

Dylan Scandinaro, who led the preparedness team and was recruited from Anthropic in February, will now concentrate on the implications of "recursive self-improving" AI. Recursive self-improvement — systems that autonomously modify or enhance their own capabilities — represents one of the most consequential and least tractable risk surfaces in frontier AI development. Whether focusing a single researcher on this area constitutes adequate institutional attention, versus the cross-functional team that previously existed, is a question the reorganization does not answer. The security concerns that arise when capable AI systems operate with greater autonomy are already surfacing in adjacent contexts; the public Sentry key AI coding agent hijack case illustrates how quickly agentic systems can become attack surfaces when oversight is sparse.

What Developers Should Expect

For teams building on OpenAI's models, pre-release risk evaluation will now be less visible and less centralized. Domain-specific teams may produce domain-specific evaluations, but the unified preparedness framework that previously aggregated those assessments into a model-level risk picture is gone. OpenAI is heading toward an IPO expected to be massive, and the organizational changes — across AGI readiness, superalignment, ethics leadership, and now preparedness — form a consistent pattern under that pressure. Safety evaluation at the frontier is becoming harder to observe from the outside precisely as the models being evaluated grow more capable.