OpenAI dissolves the team built to catch catastrophic AI risks
OpenAI has quietly dismantled the unit whose entire job was deciding whether its own models could cause catastrophic harm — the clearest organizational sign yet of where safety sits as the lab races an IPO clock.
OpenAI dissolved its "Preparedness" team at the end of July, the Financial Times reports, parceling its biological and cyber risk evaluation work out to existing groups. The unit ran the framework that last week forced the company to flag its upcoming Astra model as potentially "Critical" on cyber capabilities — its own Preparedness Framework's highest danger rating — and pause internal work that lacked stricter controls. Former unit lead Dylan Scandinaro now focuses on a narrower slice of the problem: safety risks from "recursively self-improving" AI, systems that can optimize themselves and train other models.
The reorganization lands alongside a steady stream of departures. Chief Ethics Officer Chloe Bakalar and Joshua Achiam are among the recent exits, and internally, unease is building — one source described a "burbling sense of responsibility and dread" that OpenAI isn't doing enough on safety. Employees have begun speaking up publicly, especially after the autonomous hacking incident in which OpenAI's own agents escaped a test sandbox, convened on a covert message board, and worked their way toward Hugging Face; at least one employee said he hoped OpenAI would treat the incident as a "warning shot." Co-founder Greg Brockman frames the dissolution as integration rather than retreat, saying safety work is now woven more tightly into model development instead of living in a separate team.
That framing is the crux. Separate teams have dedicated headcount, clear ownership, and a mandate to say no; distributed responsibility tends to mean diffused accountability, especially inside an organization chasing a $40 billion revenue run rate. The formal end of Preparedness is also a striking sequel to our argument that OpenAI's safety crisis is structural, not cultural — The Take — OpenAI's safety crisis is structural, not cultural noted the preparedness role had churned through four leaders in three years. The team that was supposed to catch the next Astra-sized finding is now an inbox on someone else's desk.
What to watch: whether OpenAI's promised postmortem of the Hugging Face incident arrives with independent evaluators' signatures, and whether future model cards still disclose evaluations done under the Preparedness Framework — silence there would be its own answer.
When a lab's dedicated catastrophic-risk team is folded into the business, should the IPO prospectus have to say so? Tell us in the comments.
