OpenAI Scatters Its Catastrophic Risk Team Into Product Groups

OpenAI shut down its Preparedness team at the end of July 2026, the unit charged with spotting whether its frontier models could trigger severe or catastrophic harms. The Financial Times first reported the move, citing internal sources. Biological and cyber risk work has been handed to existing groups. Former lead Dylan Scandinaro now concentrates on risks from recursively self-improving systems.

Co-founder Greg Brockman described the shift as weaving safety more tightly into everyday model development. The change arrives while the company faces a wave of safety and ethics departures and prepares what many expect will be a large IPO.

The Quiet July Shutdown

The team stopped operating as a standalone group late last month. Responsibilities for specific domains such as biosecurity and cybersecurity moved into product-aligned teams. Scandinaro, hired from Anthropic in February 2026 to head Preparedness, shifted focus to the implications of AI systems that can improve themselves and train successors.

OpenAI has not issued a lengthy public statement. Leadership frames the reorganization as efficiency and deeper integration rather than retreat. Internally, some staff describe a growing unease that the company is not doing enough on the hardest risks.

• End of July 2026: Preparedness team disbanded
• February 2026: Dylan Scandinaro hired from Anthropic as head
• October 2023: Team originally launched under Aleksander Madry
• Recent exits: Ethics lead, safety systems head, chief futurist

The Preparedness Framework itself remains in place under the Safety Advisory Group, according to accounts of the internal briefings.

That split matters. The framework still sets the scoring language. The people who once applied it full time no longer sit in one room. Product groups now carry both the release calendar and the residual risk judgments that once belonged to a separate unit.

Scandinaro’s tenure at the head of the team lasted only months. He arrived in February 2026 and saw the group closed by the end of July. The short window leaves little public record of what the unit produced under his leadership before the redistribution.

What the Team Was Built to Catch

OpenAI created Preparedness in late 2023 to close a clear gap. Frontier models were improving fast. The company needed a group that would tightly link capability tests, evaluations, and red-teaming specifically for the worst outcomes.

The original mandate covered tracking, evaluating, forecasting, and protecting against catastrophic risks spanning cybersecurity and CBRN threats, plus individualized persuasion and autonomous replication and adaptation. It also owned development of a Risk-Informed Development Policy that would set evaluation gates and protective actions before deployment.

A public challenge invited outside researchers to surface less-obvious failure modes. Winners received API credits. The exercise fed the early OpenAI Preparedness Framework beta document that still guides scoring of model capabilities against those risk bands.

Aleksander Madry led the unit at launch. Leadership later changed hands more than once. By early 2026 Scandinaro held the role with a compensation package reported as high as $555,000 plus equity.

The four risk categories were never meant to be handled as side tasks. Each one demands specialized evaluation design, threat modeling, and escalation paths that differ from ordinary product testing. Folding them into shipping teams changes who holds the pen when a capability score approaches a gate.

The Risk-Informed Development Policy was the operational hinge. It was supposed to turn framework scores into concrete go or no-go decisions before deployment. Without a dedicated owner, that hinge now depends on product managers and existing safety staff sharing the same calendar and the same release pressure.

Third Safety Unit to Go

This is not the first dedicated safety structure OpenAI has dissolved. The pattern is now clear enough that outsiders track it as a cycle.

1. May 2024: Superalignment team disbanded after co-founder Ilya Sutskever and Jan Leike left. The group had been promised 20 percent of compute to solve control of superintelligent systems.
2. 2024: AGI Readiness team also wound down; its head later departed.
3. End of July 2026: Preparedness team closed and its domain work redistributed.

Each time the public explanation stressed better integration or new priorities. Each time the dedicated long-horizon risk capacity shrank. Mission Alignment work was similarly restructured earlier. Critics inside and outside the company now treat the sequence as evidence that safety organizations function as temporary cost centers once commercial timelines tighten.

Three closures in roughly two years form a record that is hard to read as coincidence. Superalignment lost its compute promise when its leaders left. AGI Readiness followed. Preparedness is the latest. The common thread is that ring-fenced capacity for distant or catastrophic outcomes does not survive contact with accelerating product plans.

Mission Alignment’s earlier restructuring fits the same arc. Each move can be defended on efficiency grounds. Taken together, they reduce the number of internal groups whose sole job is to slow a release when the risk case is incomplete.

Integration Claim Meets Exit Wave

Brockman’s line that safety is now woven more tightly into development is the official frame. Product teams will own the bio and cyber evaluations that once sat with a specialized unit. In theory that puts the people shipping models in direct contact with the hardest failure modes.

The timing undercuts the reassurance for many observers. Chloe Bakalar, OpenAI’s only dedicated ethicist departure last month, left after less than a year. She had joined from Meta in August 2025 to work on model development ethics, human-AI interaction, and questions of machine consciousness. Johannes Heidecke, head of safety systems, and Joshua Achiam, chief futurist and former mission alignment lead, also exited recently.

Jan Leike’s 2024 resignation statement still circulates. He wrote that safety culture and processes had taken a back seat to shiny products. That phrase reappears in coverage of the latest round of departures.

safety culture and processes have taken a back seat to shiny products

Leike said that after leaving the Superalignment effort. The line has become shorthand for a longer-running tension between research-led caution and product velocity.

On X, the dominant reaction treats the July decision as confirmation of the same pressure. Posts note the irony of dissolving a team built to catch rogue-model risks while the company races toward an IPO and ever-stronger systems. One high-engagement alert simply stated the facts and let the timing speak. Crowd analysis repeatedly returns to the same observation: dedicated risk units keep proving disposable once capital-structure and competitive clocks start ticking.

The exits cluster around roles that once translated abstract risk into process. An ethics lead, a safety systems head, and a chief futurist who had led mission alignment all left in the same broad window as the Preparedness closure. That concentration thins the bench of people who could challenge a release from outside the product chain of command.

How the Framework Survives Without Its Team

The Preparedness Framework remains under the Safety Advisory Group. That continuity is the main rebuttal to claims of full retreat. Scoring bands, evaluation language, and the beta document’s structure still exist on paper and in internal process maps.

What changes is ownership of the work that feeds those scores. Biosecurity and cybersecurity evaluations now sit with product-aligned groups. Those groups already balance feature scope, latency targets, and launch dates. Adding catastrophic-risk gates to the same queue alters incentives even if the written thresholds stay constant.

The original public challenge and the framework beta assumed a unit that could absorb outside findings and turn them into protected evaluation time. Without that unit, external researchers and internal red-teamers must route concerns through teams whose primary metric is shipping. The framework can still record a high risk score. The question is how often that score still triggers delay or heavier mitigation when the owners also own the release.

Scandinaro’s narrower brief on recursive self-improvement keeps one catastrophic pathway under focused attention. The other three original categories now compete for time inside broader organizations. That is a structural trade, not a claim that the residual risks have shrunk.

Recursive Focus and the IPO Clock

Scandinaro’s new brief on recursively self-improving AI is not a soft landing. Systems that can optimize their own training and spawn improved versions sit near the top of many catastrophic-risk lists. Concentrating a strong researcher on that problem may produce sharper work than a broad team stretched across four risk domains.

Yet the broader organizational signal travels farther. OpenAI is simultaneously pursuing a push for ever-faster model inference speeds and preparing market narratives that favor rapid capability delivery. Folding the general catastrophic-risk function into product groups removes one independent voice that could slow a release or demand heavier mitigations.

Regulators and rival labs watch the pattern. Earlier voluntary commitments and summit contributions leaned on the existence of Preparedness-style capacity. When that capacity is redistributed, outside observers have less clear evidence that the hardest evaluations still receive protected attention and compute.

Internal sources quoted across reports describe a “burbling sense of responsibility and dread.” Employees who remain continue the day-to-day safety work. The question they face is whether the new structure can still surface and escalate a truly novel catastrophic pathway before it ships.

An IPO path rewards clean stories about growth and technical lead. Independent safety units that can publicly or internally stall a launch complicate that story. The July move reduces one source of such friction just as the company prepares market narratives around speed and scale.

• Framework and Safety Advisory Group retained as process backbone
• Bio and cyber evaluations moved into product-aligned teams
• Recursive self-improvement kept as a focused research brief
• Inference speed and capability delivery pushed in parallel

Why Outside Observers Read the Pattern Hard

Regulators who accepted voluntary commitments built around Preparedness-style capacity now face a thinner org chart. Summit contributions that once pointed to a dedicated catastrophic-risk team must be re-read against product groups that hold the same work alongside launch goals.

Rival labs gain a clearer competitive map. They can keep ring-fenced teams and sell that choice to enterprise buyers and policymakers, or they can match OpenAI’s leaner model and argue that integration is the industry norm. Either response reallocates scarce alignment talent across the sector.

The irony noted on X is structural rather than personal. A team hired to catch rogue-model and catastrophic pathways closed while the company races toward stronger systems and a large IPO. The facts of timing do the persuasive work without much editorial gloss.

None of this proves that residual evaluations will fail. It does show that the independent escalation path is shorter than it was in late 2023, when Preparedness launched under Madry to close a recognized gap between fast capability gains and slow risk process.

Who Feels the Shift First

Talent markets notice first. Safety researchers deciding where to spend the next five years now see a repeated cycle: build a specialized team, publish a framework, then watch it get absorbed or closed when the product roadmap accelerates. Some leave for startups or rivals that still maintain ring-fenced risk groups. Others stay and try to make the integrated model work.

Downstream developers and enterprise customers inherit the practical consequences. If bio and cyber evaluations now live inside the same organizations that own shipping schedules, the threshold for blocking a release may rise. That could mean faster feature velocity. It could also mean thinner documentation of residual risk when a new capability lands in production.

Competitors face a strategic choice. Match the leaner structure and claim equal integration, or keep dedicated teams as a differentiator for customers and regulators who still value independent catastrophic-risk capacity. Either path reshapes how the industry allocates scarce alignment talent.

The Preparedness name is gone. The risks it was created to measure have not diminished. OpenAI is betting that product teams plus a remaining framework and advisory group can carry the load. The next major model release will supply the first real test of that bet.

Until that release lands, the public record holds a framework without its original team, a narrowed recursive-risk brief, and a string of safety exits beside an IPO clock. Those pieces will shape how every subsequent safety claim from the company is weighed.