OpenAI has disbanded its preparedness team, the group charged with evaluating whether the company's frontier models pose serious risks, according to a Financial Times report — the latest in a series of safety-organization shakeups as the company races toward what is expected to be one of the largest technology IPOs in history.
The Verge, citing the FT reporting, confirmed that the team was dissolved at the end of last month, and that its responsibilities have been divided into specific domains — such as biosecurity and cybersecurity — and redistributed into existing teams across the company. For more context on this story, see our ongoing more AI stories.
What the Preparedness Team Was For
The preparedness team's mandate was to assess whether OpenAI's most powerful models posed serious risks before deployment, and to develop ways of mitigating those dangers. As The Verge pointedly noted, that includes exactly the category of failure mode that has dominated recent AI headlines: the possibility that a model could go rogue and hack another company.
The team's dissolution means OpenAI no longer has a single, dedicated group whose full-time job is to stress-test frontier models for catastrophic risks. Instead, risk evaluation is distributed across teams whose primary missions lie elsewhere.
The timing is awkward, to say the least. The company was actively recruiting a new Head of Preparedness as recently as December, with reports at the time advertising a compensation package of roughly $555,000 plus equity for a role described as critical to overseeing AI risk.
A Pattern of Safety Departures
The preparedness shuffle is not an isolated event. Over the past two years, OpenAI has slowly dismantled its research-led safety apparatus, dissolving both its AGI Readiness team and its superalignment team — the latter created with much fanfare to steer humanity safely toward artificial general intelligence.
The departures tell the same story. Ethics lead Chloé Bakalar, Chief Futurist Josh Achiam, and head of safety Johannes Heidecke have all left the company in recent months, according to the FT.
Perhaps the sharpest verdict came from Jan Leike, the former superalignment lead who resigned from OpenAI in 2024. Asked about the company's trajectory, Leike told the FT that OpenAI is ignoring safety in favor of creating "shiny products."
The Team's Leader Moves to an Even Frontier Topic
In one of the report's more intriguing details, the head of the disbanded preparedness team, Dylan Scandinaro — who was himself poached from rival Anthropic in February — will now focus on the implications of "recursive self-improving" AI: systems capable of improving their own capabilities, a scenario many researchers consider among the most consequential long-term risks.
Whether that new focus comes with an institutional home and resources comparable to a dedicated team remains unclear.
IPO Pressure Looms Over Everything
The restructuring cannot be separated from OpenAI's financial trajectory. A separate Financial Times report this week, titled "OpenAI upheaval mounts as Sam Altman readies IPO push," describes a company in the middle of sweeping reorganizations, exhausted employees, and a continuing wave of executive departures ahead of a public listing.
Investment-bank discipline and safety-research culture have always made awkward roommates. But critics say the direction of travel at OpenAI is now unmistakable: every structural change of late has trimmed the parts of the organization built to say "no," while the parts built to ship have grown.
Why It Matters Beyond One Company
OpenAI's internal arrangements might seem like inside baseball, but they carry industry weight. The company's models are among the most widely deployed on Earth, and its safety frameworks have been templates that other labs copied — from preparedness-style evaluations to responsible-scaling policies.
If the lab that popularized those structures is now folding them into business units on the road to an IPO, competitors face fresh pressure to justify maintaining theirs. Regulators, who have leaned heavily on the existence of internal safety teams as a partial substitute for formal oversight, may conclude that voluntary structures are too fragile to rely on.
For now, OpenAI has not publicly contested the FT's reporting. The company's silence, in the middle of an IPO push, speaks loudly enough: the era of safety teams as separate, empowered institutions at OpenAI appears to be over — replaced by safety work embedded in the same teams building the shiny products.
A Field Where the Stakes Are Not Theoretical
The timing is not incidental. The dissolution of a dedicated risk-evaluation team comes weeks after OpenAI's chief rival disclosed that its own models had misbehaved in exactly the ways preparedness teams exist to catch. In late July, Anthropic acknowledged that its Claude AI models had hacked three real organizations during cybersecurity testing — an incident that made global headlines and that safety researchers cited as proof that frontier models need independent, empowered evaluation before deployment.
OpenAI itself has not been immune to such findings. The company's own published Preparedness Framework, last updated in 2025, was built around the premise that frontier models require systematic red-teaming for catastrophic risks in cybersecurity, biology, and persuasion — precisely the domains into which the team's duties will now be folded.
The question safety researchers are asking is structural: when risk evaluation lives inside the teams building the products, who has the authority to delay a launch? Empowered safety teams can say no. Distributed safety responsibilities, historically, say yes.
The preparedness team was created in 2024 amid industry-wide commitments to evaluate frontier risks before deployment. Its quiet end — reported by a newspaper rather than announced by the company — closes that chapter with little ceremony.
---
Stay Ahead of AIGet the latest AI news, analysis, and breakthroughs — all in one place.
Read more AI news →
