OpenAI disbanded its Preparedness Team at the end of July, according to the Financial Times, ending the standalone group that had been tasked with evaluating major and catastrophic risks tied to frontier AI models. The work itself has not been dropped. Instead, responsibilities once concentrated in that team are being reassigned to senior staff inside existing research groups, including teams covering biosecurity and cybersecurity.
What the Preparedness Team was set up to do
OpenAI formally created the Preparedness Team in October 2023. It was initially led by MIT professor Aleksander Madry. The group was not responsible for routine ChatGPT content moderation. Its job was to examine more extreme frontier AI risks, including whether a model had developed capabilities that could cause serious societal harm.
OpenAI had identified several key research areas for the team:
- Cybersecurity attacks
- Chemical, biological, radiological and nuclear risk, or CBRN
- Large-scale persuasion and manipulation
- AI self-replication and adaptation
The team also carried out capability testing and red teaming for models that had not yet been released, with the aim of judging whether those systems had crossed dangerous thresholds. In 2025, OpenAI updated its Preparedness Framework and added a Capabilities Report and a Safeguards Report.
Responsibilities are being split across specialized lines
The Financial Times, citing multiple people familiar with the matter, reported that OpenAI did not shut down this category of work after dissolving the Preparedness Team at the end of July. The company shifted specific risk domains to senior employees inside its current structure, with biosecurity and cybersecurity handled by their respective teams.
Dylan Scandinaro, who had led the Preparedness Team, has moved to another safety problem at the frontier: Recursive Self-Improvement. The article described that line of work as AI that does more than complete tasks and starts helping improve its own capabilities, potentially even training the next generation of models.
OpenAI co-founder Greg Brockman said the organizational change reflects the need for 「stronger safety measures」 as model technology advances quickly. He said the company is integrating research, safety and security work more deeply into the model development process.
Part of a longer run of safety-team restructurings
The dissolution of the Preparedness Team is not the first time OpenAI has removed an independent AI safety unit.
In 2024, the company disbanded the Superalignment Team led by co-founder Ilya Sutskever and Jan Leike. That group had focused on how to control future superintelligent systems that could become smarter than humans. When Leike left, he publicly criticized OpenAI, saying the company’s 「safety culture and processes have taken a backseat to shiny products」.
After that, teams including AGI Readiness and Mission Alignment were also reorganized or dissolved. OpenAI’s position has been that safety should no longer be treated as a separate final-stage check, but should be folded directly into research and product development.
In July this year, Johannes Heidecke, head of OpenAI Safety Systems, also announced his departure. OpenAI then placed its safety teams under Mia Glaese, vice president for research and safety. At the time, chief research officer Mark Chen said model training is moving faster and release cycles are getting shorter, so safety work needs to enter research and product decisions earlier.