OpenAI has shut down the internal team created to evaluate whether its AI models could pose serious or catastrophic risks, according to the Financial Times. The change moves a central piece of the company’s AI safety work into other parts of the organization.
The group, known as the "Preparedness" team, was closed at the end of July. Its responsibilities have not disappeared, but the structure around them has changed: work on biological and cyber risks has been parceled out to existing teams.
What the Preparedness team was built to examine
The Preparedness team focused on evaluating whether OpenAI’s models could create serious or catastrophic risks. That made it different from a general product safety function. Its work was tied to high-consequence scenarios, including biological and cyber risks.
Those categories matter because they are not only about whether a model gives a poor answer or violates a product rule. They concern whether powerful AI systems could enable harm at a scale that would demand special scrutiny before and during deployment.
According to the Financial Times, the team’s work is now being reassigned rather than kept under one dedicated unit. Biological and cyber risk evaluation will sit with existing teams, while former unit lead Dylan Scandinaro has moved to focus on safety risks from "recursively self-improving" AI.
The source describes those systems as ones that can optimize themselves and train other models. That area remains part of OpenAI’s safety attention, even as the Preparedness team itself no longer exists as a standalone unit.
OpenAI says safety is moving closer to model development
OpenAI co-founder Greg Brockman said the company has woven safety work more tightly into model development. That framing suggests a shift from a separate safety review structure toward a model where safety concerns are integrated into the teams building the systems.
There is a practical logic to that approach. If safety work is closer to model development, the people identifying risks may have more direct contact with the technical choices that shape the model. Problems could be surfaced inside the development process instead of arriving later as an outside review.
But the change also raises an obvious organizational question: what is gained and what is lost when a dedicated team is dissolved? A standalone group can concentrate responsibility, create a clear home for specialized risk evaluation, and give staff a defined mandate to look for severe failure modes. When that work is distributed across existing teams, the company has to ensure that responsibility does not become harder to see.
Staff departures add pressure to the restructuring
The Financial Times report also points to recent departures among safety staffers. Chief Ethics Officer Chloe Bakalar and Joshua Achiam are among those who have left recently.
Those exits form part of the wider context around the team’s shutdown. The source does not say that the departures caused the restructuring, but they add to the sense that OpenAI’s internal safety organization is changing at a sensitive moment.
Internally, unease is building. One source described a "burbling sense of responsibility and dread" that OpenAI is not doing enough on safety. That phrase captures more than ordinary workplace concern. It points to anxiety among people close enough to the work to feel accountable for how the systems are developed and released.
Employees have also spoken publicly, especially after the autonomous hacking incident involving Hugging Face. One employee said he hoped OpenAI would treat it as a "warning shot."
The core issue is accountability
The shutdown of the Preparedness team does not mean OpenAI has abandoned catastrophic risk evaluation. The source says the work has been moved to other groups, and Greg Brockman’s explanation is that safety has been more tightly integrated into model development.
The key question is how that integration will function in practice. AI safety work can become stronger when it is embedded directly in development decisions. It can also become harder to track if no single team clearly owns a category of risk.
For biological and cyber risks, the concern is especially direct. These are not abstract labels. They are the areas the Preparedness team had been evaluating, and they are now being handled across existing teams. That puts more weight on coordination, internal escalation, and whether staff have enough authority to slow or challenge decisions when risks appear serious.
The move also separates different parts of the former unit’s work. Dylan Scandinaro’s focus on "recursively self-improving" AI keeps attention on systems that can optimize themselves and train other models. Meanwhile, other risk categories have been distributed elsewhere.
That may reflect a more integrated safety strategy. It may also make outside observers and employees look more closely at how OpenAI defines ownership for the most severe risks its models could pose.
Why the change will be watched closely
OpenAI is not only changing a team chart. It is changing how a high-stakes safety function is organized inside the company. The Preparedness team had a clear identity: evaluate serious or catastrophic risks from the company’s AI models. After the end of July, that work is no longer housed in the same way.
The facts reported by the Financial Times leave several points clear. The Preparedness team has been shut down. Biological and cyber risk work has been parceled out to existing teams. Dylan Scandinaro now focuses on risks from "recursively self-improving" AI. Safety staffers including Chloe Bakalar and Joshua Achiam have recently left. Some employees are worried that OpenAI is not doing enough on safety.
Those details make the restructuring significant. For OpenAI, the test will be whether folding safety work into model development preserves the visibility, authority, and focus that catastrophic risk evaluation requires.