OpenAI dissolves the team that judged whether a model was too dangerous to ship

OpenAI disbanded its preparedness team at the end of last month, the Financial Times reported, as summarised by The Verge ^[1]. That team's job was to assess whether models posed serious risks and to develop ways to mitigate them ^[2], and according to the FT its responsibilities have been split by area, into buckets such as bio and cyber, and moved into existing teams ^[3].

Read the second half of that sentence slowly, because it is the part that changes anyone's planning. A function that existed as its own group now sits inside the groups that build and ship ^[9]. The FT's account does not describe a new gate replacing the old one; it describes the same questions being asked by people whose day job is the product ^[3].

The Verge frames this as the latest in a sequence rather than a one-off: over the past few years OpenAI has also dissolved its AGI readiness and superalignment teams ^[6], which makes preparedness the third safety-oriented group to be taken apart ^[10]. Ethics lead Chloe Bakalar, chief futurist Josh Achiam and head of safety Johannes Heidecke have all left recently ^[5]. Jan Leike, who resigned from OpenAI in 2024, told the FT that the company was ignoring safety in favour of creating "shiny products" ^[7]. Dylan Scandinaro, who led the preparedness team and was hired away from Anthropic in February, will now work on the implications of "recursive self-improving" AI ^[8]. All of this is happening, per The Verge, as the company moves toward an expected large IPO ^[4].

For operators, the practical question is narrow. If your model risk assessment, your vendor questionnaire, or your regulator-facing documentation leaned on the existence of an independent group at OpenAI that could decline to release a capability, that load-bearing assumption is now a claim you cannot make ^[1][3]. Nobody has said the assessments stopped. But an assessment owned by a shipping team is not the same control as an assessment owned by a group with no shipping deadline, and the difference matters most in exactly the cases you were relying on it for: a capability that is commercially attractive and operationally awkward.

Rebuilding in-house is unglamorous work. It means your own red-team budget for the capabilities you actually expose, your own abuse monitoring on your own traffic, contract language that gives you notice and rollback rather than trusting release notes, and an internal owner who is allowed to say no to a launch. If you were treating the vendor's internal process as a compensating control in an audit, that mapping needs to be redone now, not at renewal.

What to watch: whether the new bio and cyber owners publish anything under their own names, so that external parties can tell who is accountable for a given judgment ^[3]; whether Scandinaro's recursive self-improvement remit produces published criteria or stays internal ^[8]; and whether IPO-related disclosure forces OpenAI to describe its safety governance in language that can be held against it ^[4]. Further senior departures would be the cheapest signal to track ^[5].