The Financial Times reported on Aug. 16 that OpenAI disbanded its Preparedness team, the unit charged with assessing catastrophic risks from AI models, at the end of July, folding its responsibilities into other departments. The report arrived as a stretch of senior departures was thinning the company’s executive ranks. Engadget, updating the story on Aug. 18, said OpenAI denies the team was disbanded, describing the change as a reorganization.
The semantic dispute is the story. The Preparedness team was created in late 2023 to test models for dangers ranging from cyberattacks to biological threats, with a mandate to keep capabilities judged too dangerous out of released products. Whether its functions were dissolved and scattered, or reorganized and absorbed, the practical effect described in both versions is the same: there is no longer a dedicated unit with the team’s name, its staff, or its singular focus.
The timing compounds the sensitivity. OpenAI has filed confidentially for an initial public offering and is widely expected to reach public markets soon after Anthropic, its chief rival, and the composition of its safety apparatus is exactly the kind of thing that regulators, investors, and competitors will examine in the prospectus. A company preparing to sell shares to the public is not usually eager to explain, in the middle of the roadshow preparation, why its catastrophic-risk assessment unit no longer exists.
The company’s explanation, according to the Engadget update, is that safety work continues inside other teams and that the change reflects how the organization has matured. OpenAI has said the reorganized structure keeps safety evaluations attached to the teams that build and deploy models, rather than in a separate group that could be seen as an afterthought. Defenders of the change argue that safety was becoming an isolated function and that embedding it in development is the stronger arrangement.
Critics read the same facts differently. The Preparedness team was founded with a public commitment to independence, with researchers empowered to flag risks even when the commercial pressure to ship was at its highest. Scattering its functions into departments that report through product and engineering chains removes that independence, they argue, and the wave of senior departures suggests the company’s culture of caution is being traded for speed as the IPO approaches.
The disagreement itself carries information for the market. If OpenAI’s account is accurate, the company has reorganized safety in a way it believes is stronger, and the change is manageable. If the Financial Times’ account is accurate, a company racing to list has quietly dismantled the unit that was supposed to be its conscience, and the risk disclosures in the prospectus will need careful reading. Investors cannot know which version is true from outside, and that uncertainty is itself a risk.
Regulators are watching the same dispute. The European Union’s AI Act imposes binding obligations on the most powerful models, including assessments of systemic risk, and the details of how those assessments are staffed and structured will matter for compliance. A company that cannot clearly explain who evaluates catastrophic risk, and with what independence, will have a harder conversation with every regulator it meets.
The broader industry is watching too. Every frontier lab has a version of this tension, between the teams that build capabilities and the teams that are supposed to limit them, and the balance of power between the two has shifted repeatedly as the commercial stakes have risen. How OpenAI resolves this particular dispute, and what its prospectus says about it, will be read as a signal about how the entire industry is managing the same balance.
For now, the question is unresolved in public. OpenAI says the team was reorganized; the Financial Times says it was disbanded. Both versions agree on the underlying fact that the dedicated unit no longer exists, and both versions will be tested when the company’s filings and its future safety disclosures tell investors which account the company will live by.
The dispute echoes a pattern that has followed OpenAI since its founding. The company’s original structure, built around a nonprofit mission to develop AI safely, has given way to a commercial organization racing to scale, and every phase of that transition has been accompanied by arguments about whether safety work was being diluted. The Preparedness team itself was created in the aftermath of a leadership upheaval in 2023, when the board fired and then reinstated the chief executive, and its founding mandate was a public promise that catastrophic risk would have an independent voice inside the company.
The history matters to investors because it is a history of reorganizations. OpenAI has repeatedly restructured its safety functions under commercial pressure, and each restructuring has been followed by departures of researchers who said the changes weakened oversight. The current dispute fits the pattern, and the question is whether this reorganization is the one that finally lands in a prospectus as a liability.
The stakes for the industry are concrete. Every frontier lab that files to go public will face the same scrutiny, and the way OpenAI accounts for its safety structure will set the template for how the market evaluates the others. If the prospectus describes a safety function that is embedded and accountable, investors may accept the reorganization; if it reads like a dissolution, the company will spend its first months as a public company defending the decision.
