IA · 17 August 2026 · 4 min read
OpenAI Disbands Its Preparedness Team, Shifting Frontier AI Safety into Product Units
In brief: OpenAI has confirmed the disbanding of its Preparedness division, the dedicated team responsible for evaluating and mitigating catastrophic risks in frontier models, including autonomous cyber threats and biological misuse. These oversight responsibilities are being divided and integrated directly into existing engineering and product teams, reflecting the company's broader transition toward a commercial, product-first structure ahead of its anticipated public offering.
by Team Mocchi's
A Structural Shift in Frontier AI Safety
OpenAI has officially disbanded its Preparedness team, the internal unit established to evaluate, stress-test, and mitigate catastrophic risks posed by cutting-edge artificial intelligence models. As reported by The Verge, citing Financial Times reporting, the restructuring took effect late last month, marking a significant change in how the company organizes its safety oversight.
The Preparedness team was tasked with conducting systematic evaluations prior to the public deployment of frontier models, setting measurable thresholds for extreme hazards such as automated cyberattacks, synthetic biology threats, and unauthorized autonomous behavior. Following the dissolution of the standalone team, oversight responsibilities will be split into domain-specific tracks—such as cybersecurity and biosecurity—and embedded directly within existing product engineering groups.
Decentralized Governance and Product-Led Alignment
This organizational pivot also alters the focus of key technical leaders. Dylan Scandinaro, who led the Preparedness team after joining from Anthropic earlier this year, remains at OpenAI to focus on the long-term implications of recursive self-improving AI architectures.
While decentralizing safety functions aligns with efforts to streamline product cycles and integrate safeguards directly into feature development, industry observers have raised questions regarding the independence of these evaluations. Embedding safety audits within commercial development units introduces potential friction, as engineering teams facing tight release schedules and market competition may experience competing incentives when balancing deployment timelines against conservative risk thresholds.
Research Departures and the Road to an IPO
The dissolution of Preparedness is the latest in a series of reorganizations affecting OpenAI's dedicated safety and alignment groups over the past several quarters. The company previously dissolved its Superalignment team and its AGI readiness unit, shifting away from standalone frontier research structures toward operational engineering.
This shift has coincided with the departure of several prominent safety and governance figures, including ethics lead Chloé Bakalar, chief futurist Josh Achiam, and head of safety Johannes Heidecke. Former researchers, such as Jan Leike, have openly expressed concern that rigorous, centralized safety oversight is increasingly taking a backseat to rapid commercial product releases—a transformation that comes into sharp focus as OpenAI moves toward a massive public listing.
Mocchi's take
For enterprise teams and European businesses building software on top of frontier AI APIs, this governance shift underscores a crucial reality: safety and robust guardrails cannot simply be outsourced to foundation model providers. As AI labs embed compliance into commercial product cycles to accelerate release velocity, the burden of red-teaming, output validation, and sandboxing shifts decisively downstream to software integrators. Engineering teams must design their own independent verification layers, ensuring domain-specific compliance rather than assuming the underlying base models are inherently safeguarded against operational risks.