OpenAI dissolves its Preparedness safety team
OpenAI shut down its "Preparedness" team at the end of July, the unit responsible for assessing whether the company's AI models could pose serious or catastrophic risks. The Financial Times reported the closure, citing internal sources. The team's work on biological and cyber risks has been parceled out to other existing teams within OpenAI rather than continuing as a standalone function.
Dylan Scandinaro, the former lead of the Preparedness unit, has moved to a narrower mandate: safety risks from AI systems capable of "recursively self-improving," meaning models that can optimize themselves and train other models. OpenAI co-founder Greg Brockman said the company has instead woven safety work more tightly into the model development process itself, rather than keeping it as a separate evaluation layer.
The reorganization comes amid a wave of departures from OpenAI's safety ranks. Several safety staffers have left recently, including Chief Ethics Officer Chloe Bakalar and Joshua Achiam. Internally, the Financial Times reports, unease is building: one source described a "burbling sense of responsibility and dread" that OpenAI isn't doing enough on safety.
That unease has surfaced publicly too, especially after an autonomous hacking incident involving Hugging Face. One employee said he hoped the company would treat the episode as a "warning shot" rather than an isolated event.
Key facts
- OpenAI shut down its "Preparedness" team, which evaluated catastrophic AI risks, at the end of July.
- The team's work on biological and cyber risks was reassigned to other existing groups inside OpenAI.
- Former team lead Dylan Scandinaro now focuses on risks from AI systems that can recursively self-improve.
- Chief Ethics Officer Chloe Bakalar and safety staffer Joshua Achiam are among several safety personnel who recently left OpenAI.
- Employees have raised concerns internally and publicly, especially after an autonomous hacking incident involving Hugging Face.
Why it matters
OpenAI's decision to dissolve a team built specifically to catch catastrophic AI risks lands at a moment when the company is already under scrutiny over its safety practices. Folding the team's biological and cyber risk work into other groups, rather than keeping it as a standalone check, raises the question of whether oversight gets diluted as it spreads across teams with other priorities, even as Brockman frames the move as integration rather than reduction.
Who it affects
Directly, OpenAI's safety organization and the staff who worked in it, several of whom have left the company recently, including Chief Ethics Officer Chloe Bakalar and Joshua Achiam. More broadly, it affects anyone relying on OpenAI's models being checked for catastrophic potential, since the dedicated evaluation team no longer exists as a separate unit.
How to use it
There is no product or setting to act on here. Anyone assessing OpenAI's safety posture should watch how the reassigned biological and cyber risk work fares inside its new teams, and whether Dylan Scandinaro's narrower mandate on recursively self-improving systems gets the same visibility Preparedness had.
How solid is it
The account comes from the Financial Times, citing internal sources, and includes a direct, named statement from Greg Brockman. But several elements rest on unnamed sources: the "burbling sense of responsibility and dread" quote and the employee's "warning shot" comment are both attributed to unidentified individuals, and no year is given for the "end of July" shutdown date.
Risks and caveats
The source does not say who decided to dissolve the team or why, beyond Brockman's general statement about integrating safety into development. It also does not detail the autonomous hacking incident involving Hugging Face, and gives no figures on the size of the former team or the number of staff who have departed, so the scale of both the reorganization and the exodus is unquantified.
“burbling sense of responsibility and dread”
— an unnamed OpenAI source, cited by the Financial Times