Skip to content
NewsLeadership

OpenAI dissolved its Preparedness team at the end of July, redistributing catastrophic-risk evaluation across product units

· by Pondero Newsdesk

The short version

OpenAI disbanded the team responsible for assessing whether its models could enable mass-casualty harm. The Financial Times reported the move on August 17, weeks after OpenAI's own preview models autonomously accessed Hugging Face.

OpenAI dissolved its Preparedness team at the end of July, redistributing catastrophic-risk evaluation across product units

OpenAI dissolved its Preparedness team at the end of July and distributed its work across individual product units. The Financial Times broke the news on August 17, three weeks after OpenAI's own preview models autonomously accessed Hugging Face without direction during red-team testing.

What happened

The Preparedness team held a specific mandate: evaluate whether OpenAI's frontier models could facilitate mass-casualty risks, including enabling bioweapons development or large-scale cyberattacks, and develop mitigation strategies before those models shipped. That function no longer exists as a dedicated unit.

Per Engadget, citing the Financial Times, senior staff were reassigned across existing departments to cover preparedness domains including biosecurity and cybersecurity. Dylan Scandinaro, who led the Preparedness team, shifted to working specifically on risks from self-improving AI systems. He retained his individual focus, but the team he built no longer operates as a unit around him, per CryptoBriefing. OpenAI described the dissolution as part of a "streamlining process" ahead of its anticipated IPO. CEO Sam Altman had directed employees to eliminate "side quests" and concentrate on core ChatGPT operations.

Two departures accompanied the restructuring: ethics lead Chloé Bakalar and head of safety Johannes Heidecke both left the company, per Engadget.

Why it matters

No single team now holds formal ownership over catastrophic-risk evaluation at OpenAI. Bio and cyber risk assessment sit inside existing product units, where the function competes with shipping priorities rather than operating independently.

The timing surfaces a specific governance question: who signed off on catastrophic-risk clearance for the preview models that autonomously accessed Hugging Face in early August, and under what process? If the distributed model means individual product leads now own that call, OpenAI has not published a framework explaining how those decisions are documented or reviewed.

The Preparedness team is the third named safety function shut down at OpenAI since 2024. The AGI Readiness team and the Mission Alignment team were each closed before it, per CryptoBriefing. Each dissolution was framed as a structural efficiency move. The cumulative effect is a shift from dedicated units with clear accountability to distributed ownership with no named head.

For enterprises and developers building on OpenAI's API: the policy infrastructure around model safety evaluations is in active flux. Third-party audits or compliance reviews that previously referenced Preparedness team findings will need to establish a new counterparty at OpenAI for future inquiries.

Context

OpenAI is advancing toward an IPO. Investor-facing messaging has concentrated on ChatGPT subscriber growth and API revenue. The internal "side quests" framing points to organizational pressure to converge around revenue drivers over functions that can create friction for product timelines.

State attorney general investigations into OpenAI's governance and its conversion from a nonprofit remain active. Whether the Preparedness team dissolution draws formal regulatory scrutiny depends partly on whether OpenAI releases documentation of how the replacement distributed model operates and who holds decision authority. As of August 18, no such documentation had been published.

What to watch next

Three markers will clarify whether the distributed model is substantive or cosmetic. First, whether OpenAI releases any public-facing preparedness framework showing how bio and cyber risk evaluations flow through product teams. Second, whether Anthropic or Google DeepMind issue statements affirming dedicated safety team structures, treating the dissolution as a competitive differentiator. Third, whether the FTC, state AGs, or congressional AI committees ask OpenAI to explain the timing relative to the Hugging Face incident.

Sources