OpenAI's third safety team closure in two years is an IPO governance fact, not an org chart tweak

OpenAI quietly disbanded its Preparedness team at the end of July, according to the Financial Times, and redistributed its bio and cyber risk assessment across existing teams rather than replacing the unit ^[1][3]. That makes three safety-focused teams closed in roughly two years, which is the sort of pattern that ends up in a risk factor rather than an org chart footnote ^[4].

The team's mandate was narrow and specific: evaluate whether frontier models posed serious risks, including rogue systems hacking other companies or enabling biological threats, and work out how to stop those scenarios ^[2]. That is now a domain-by-domain responsibility spread across groups that have other jobs ^[3].

The sequence matters more than any single closure. The AGI Readiness team was disbanded in 2024 ^[5], the Mission Alignment team was closed in February 2026 ^[6], and Preparedness followed in July ^[1]. Three dissolutions across roughly two years works out to about one every eight months ^[7]. Each came with the same explanation, that the work is being integrated elsewhere rather than eliminated ^[8], and co-founder Greg Brockman has argued that embedding safety directly into model development and product teams strengthens it relative to siloing it in a separate group ^[9]. That argument may be correct. It is also unfalsifiable from the outside, which is the point an investor should notice: a distributed function has no separate headcount, no named owner, and no distinct artifact to inspect.

The paper apparatus is still there. OpenAI has maintained a Preparedness Framework since at least 2024, aimed at tracking and mitigating severe harms from advanced AI ^[10], and updated it as recently as April 2025 to define High and Critical capability thresholds and to introduce Safeguards Reports reviewed by an internal Safety Advisory Group ^[11]. The source material does not say who now produces those reports or owns the framework day to day. For anyone underwriting this company, that gap is the story rather than a detail.

Personnel churn runs alongside the restructuring. COO Brad Lightcap and Chief Ethics Officer Chloe Bakalar have both departed ^[12]. Dylan Scandinaro, who formerly headed Preparedness, has reportedly shifted his focus to risks from self-improving systems, but the team he built no longer exists around him ^[13]. Internal sources have flagged concerns about what repeated departures mean for continuity in risk assessment work ^[14].

The context is an expected listing described as one of the largest tech IPOs in years, and a multi-year drift from a research-led nonprofit structure toward a conventional corporate model ^[15][16]. Pre-IPO, a safety organisation is not only an ethics posture; it is a control environment, and control environments get diligenced. A named unit with a mandate and a reporting line is auditable. Domain-embedded responsibility is cheaper to run and considerably harder for an outside party to verify, which is a discount to governance quality whether or not the underlying work continues at the same intensity.

Two caveats on the evidence. This is FT reporting relayed through secondary outlets, with no headcount figures and no OpenAI statement in the material beyond the standing integration rationale ^[1][8]. And nothing here shows safety output has fallen, only that the structures which made it legible have been dismantled three times ^[4].

What to watch: whether the Preparedness Framework is revised to name its new owners, whether Safeguards Reports keep appearing under the Safety Advisory Group, and whether pre-offering disclosure specifies who signs off on a Critical threshold determination ^[11].