OpenAI has quietly dissolved its Preparedness team, the internal group charged with assessing whether the company's most advanced AI models could enable catastrophic harms, according to a Financial Times report that emerged over the weekend. The unit was wound down at the end of July, with its responsibilities for biological and cyber risk assessment split up and folded into existing teams rather than continued as a standalone group.
The Preparedness team's job was narrow but consequential: work out whether a given frontier model was capable of things like helping someone build a bioweapon, or acting as a "rogue" system that could hack into other companies' networks on its own initiative, and then figure out how to stop that from happening before the model shipped. Instead of keeping a dedicated group focused on that question, OpenAI has now distributed the work by subject area, according to a report from Digit that corroborated the FT's account, with senior staff moved into existing teams to handle bio and cyber risk specifically. OpenAI's own published preparedness framework has long organized frontier-model risk into categories including biological and chemical weapons, cyber operations, and model autonomy — the possibility that a system could act to undermine human oversight of its own operations. It's that last category, critics say, that becomes harder to monitor once the team built to track it no longer exists in its original form.
A pattern of dissolved safety teams
The Preparedness team is the third safety-focused unit OpenAI has dissolved in roughly two years. The company shut down its Superalignment team, co-led by co-founder Ilya Sutskever and researcher Jan Leike, in May 2024, shortly after both departed the company. Its AGI Readiness group was also dissolved that year. More recently, a Mission Alignment team formed in September 2024 was disbanded in February 2026. Each time, OpenAI has offered a version of the same explanation: the work isn't being abandoned, it's being absorbed into other parts of the organization.
The reshuffle has coincided with a wave of departures among OpenAI's safety and ethics staff. Ethics lead Chloé Bakalar, Chief Futurist Josh Achiam, and head of safety Johannes Heidecke have all left the company recently, according to the FT's reporting. Dylan Scandinaro, who was hired away from rival lab Anthropic in February to lead the Preparedness team, is staying on but shifting his focus to the implications of "recursive self-improving" AI — even as the team he was brought in to run no longer exists as a distinct unit around him.
Timed to a blockbuster IPO
The dissolution lands as OpenAI reorganizes ahead of what is expected to be one of the largest technology listings ever. The company confidentially filed paperwork for an initial public offering with the Securities and Exchange Commission in June, and has reportedly been weighing whether to list later this year at a valuation below $1 trillion or wait until 2027 for a cleaner shot at crossing that threshold. That looming IPO has become the backdrop against which nearly every recent OpenAI organizational change is now read, including this one.
Reaction from AI safety researchers has been sharply critical. Leike, who left OpenAI in 2024 to join Anthropic and has become one of the company's most persistent outside critics, told the Financial Times that OpenAI was ignoring safety in favor of creating "shiny products."
Ignoring safety in favor of creating "shiny products."
Jan Leike, former OpenAI Superalignment co-lead, quoted by the Financial Times
OpenAI has not issued a public statement addressing the dissolution directly, and did not comment for the FT's story. The company's public-facing preparedness framework still describes tracking, evaluating, and protecting against catastrophic AI risk as a company-wide priority, though it no longer names a single team responsible for carrying that out.
For now, the practical effect is that OpenAI is entering a period of releasing increasingly capable models — and pursuing an IPO that will invite heightened public and investor scrutiny — without the dedicated risk-evaluation unit it built specifically for that purpose two years ago. Whether the distributed model proves as rigorous as a standalone team is a question outside researchers say they'll be watching closely as the IPO process unfolds in the coming months.
