AI Safety Governance · OpenAI
Another watchdog for
catastrophic risk just disappeared.
OpenAI has dissolved "Preparedness," the team dedicated to assessing whether its models could cause catastrophic harm. The work itself was reportedly handed to existing teams — but this is the third time a dedicated oversight group has gone away. Scrutiny of the company's safety governance keeps getting sharper.
The Preparedness team
dissolved at end of July
OpenAI dissolved "Preparedness," its team dedicated to assessing catastrophic model risk, at the end of July, reassigning its biological, chemical, and cyber-risk work to existing teams, according to THE DECODER and other outlets. Details are in THE DECODER's reporting. Dylan Scandinaro, who had led the unit (and joined from Anthropic in February 2026), reportedly moved to a new role assessing risk from AI that can retrain or improve itself. Co-founder Greg Brockman has said the change reflects weaving safety work more tightly into the model-development process itself, rather than isolating it in a separate team.
Over the same weeks, several safety staffers have also departed, including Chief Ethics Officer Chloe Bakalar and Joshua Achiam. The Preparedness Framework itself reportedly continues to operate under the Safety Advisory Group — but the dedicated, independent organization that had run it is gone.
The third safety-team
breakup
May 2024: Superalignment dissolves
The team researching long-term AI risk was disbanded. Co-lead Jan Leike and others left, and reports widely cited safety culture taking a back seat to product development.
Early 2026: Mission Alignment dissolves
The team overseeing adherence to OpenAI's founding mission was disbanded, its work folded into existing divisions.
End of July 2026: Preparedness dissolves
The last dedicated team assessing catastrophic risk was disbanded — coming just days after the Astra model was flagged as a critical cyber risk on August 11.
"The work continues" isn't
the same as "the team continues"
OpenAI's explanation has stayed consistent: the work moves to existing teams, and the framework keeps running. But a dedicated, independent group and a staffer juggling this alongside other duties don't carry the same capacity for sustained monitoring, or the same independence to push back internally. In under three years, three separate dedicated oversight functions — long-term risk, mission adherence, and catastrophic risk — have each gone away in turn. Whatever the merits of any single explanation, the pattern itself keeps adding to outside doubts about how much teeth the safety governance actually has.
The work may carry on —
the watchdog role isn't guaranteed to carry the same weight.
Enterprises should weigh
this into vendor risk reviews
Business and PM teams building on OpenAI's models should treat this as one more data point for evaluating a vendor's AI governance. A framework that continues on paper isn't the same guarantee as a dedicated team running it full-time — that difference can show up directly in how fast risks get disclosed and how accountable the company is to outside scrutiny. TechBuzz's coverage also notes that the dissolution lines up with a broader leadership reshuffle. At contract-renewal or vendor-review time, it's worth adding "recent safety-team changes" as an explicit line item to check. Day-to-day personal use is barely affected.
The counter-argument
also holds up
That said, breaking up a dedicated team doesn't automatically mean safety gets worse. Brockman's framing — that folding safety evaluation into the development process itself, rather than isolating it in one team, can shorten the distance between spotting a risk and acting on it — is a real possibility, not just spin. Whether that actually plays out in practice is hard to verify from the outside with the information available today. How the Preparedness Framework performs going forward, and how much real independence the Safety Advisory Group retains, are both things worth tracking as more disclosure comes out.
None of this needs a snap judgment, but it's not a topic to shelve indefinitely either. Since similar reorganizations could show up at other major labs, it's more useful to keep comparing safety-governance structures across companies side by side over time than to take any single company's explanation at face value.