What the Preparedness team was and why it existed
The Preparedness team was OpenAI's unit with a clearly scoped mandate: assess whether a frontier model carries serious, potentially catastrophic risks - from autonomous systems compromising external infrastructure to contributions toward biological or chemical threats - and, on that basis, develop countermeasures before a model becomes publicly available or accessible through a partner program. Per reporting drawing on internal sources and the Financial Times, the team was disbanded at the end of July 2026. Instead of a standalone unit, OpenAI now distributes responsibility for bio and cyber risk assessment across existing subject-matter teams.
OpenAI has offered what's now a recurring explanation for the move: the work isn't disappearing, it's simply being integrated into other parts of the organization. Co-founder Greg Brockman is quoted taking the position that this approach actually strengthens safety, because it embeds it directly into model development and product teams rather than leaving it in a separate review body. That's a plausible organizational logic - but it can't be cleanly separated from a second, less organizational explanation: the dissolution falls in a period of intensive restructuring ahead of an expected IPO.
The third team in two years
The Preparedness dissolution doesn't stand alone. It's the third safety-focused team OpenAI has dissolved in roughly two years: the AGI Readiness team in 2024, the Mission Alignment team in February 2026, and now, in July 2026, the Preparedness team. Each of these three dissolutions came with a version of the same explanation - responsibility isn't being lost, just reorganized. A single event of this kind could be read as a legitimate organizational decision. Three structurally similar dissolutions in two years form a pattern that's harder to dismiss as mere restructuring.
Add to that pattern two prominent departures that fall close together in time. Chloé Bakalar, who served as OpenAI's only dedicated ethics chief since August 2025, left the company in July 2026 - without an official announcement from OpenAI. Her own public quote from a conference before her departure captures the difficulty of her own role: ethics is 'actually everyone's responsibility', as she put it, while well aware that exactly this diffusion of responsibility was what had made a dedicated role like hers necessary in the first place. On 11 August 2026, it was also confirmed that Brad Lightcap, longtime COO and most recently head of special projects, is leaving the company to 'start something new'.
The connection that makes this relevant for our audience
Three days ago, we wrote here about GPT-5.6-Cyber: a model that, per OpenAI's own benchmark, solves 95 percent of offensive security tasks - versus 1.5 percent for the standard model with default safeguards - accessible only through a tightly held partner program. The exact unit that would have reviewed such an internal capability before release - whether it's sufficiently contained, who gets access, which misuse scenarios are realistic - was the Preparedness team. It has not existed as a standalone unit since the end of July.
That doesn't necessarily mean GPT-5.6-Cyber itself was inadequately reviewed - the exact sequence of events between the Preparedness dissolution at the end of July and the Daybreak Red rollout in early August can't be fully reconstructed from publicly available information. But it does mean that, at the time of that rollout, the specialized review body separate from day-to-day operations for exactly this kind of capability no longer existed in its previous form. Distributing that responsibility 'by domain across existing teams' inevitably moves it closer to teams that have an interest in shipping a product.
What the independent safety index already showed beforehand
This development connects to an independent assessment that predates it: the Future of Life Institute's AI Safety Index, with data through early June 2026 and published in July 2026, graded nine leading AI labs (including Anthropic, OpenAI, Google DeepMind, Meta, xAI) across 37 indicators in six domains. None of the nine labs earned better than a C+ - Anthropic led with a C+, OpenAI received a C. Before the Preparedness dissolution, the index had already documented that several labs, including Anthropic, OpenAI, Google DeepMind, and Meta, are weakening their own earlier pledges to unilaterally pause model development at specified risk thresholds - what the reviewers themselves called 'moving the goalposts'.
Against that backdrop, the Preparedness dissolution reads not as an isolated event, but as a concrete, current example of an already-documented industry trend: safety governance at frontier AI labs is being scaled back rather than built up, even as the models' own capabilities keep increasing.
What this means in practice
For a company using OpenAI as a strategic AI vendor, or seriously considering it, this news isn't a reason for an immediate vendor switch - but it's a clear prompt not to take public safety commitments at face value as a current state of affairs.
- In your own vendor due diligence, explicitly ask which internal unit is currently responsible for assessing new, high-risk capabilities - not just whether a dedicated team existed at some point in the past.
- Use independent assessments like the Future of Life Institute's AI Safety Index as an additional information source alongside a vendor's own statements, precisely because self-reported information tends to be more favorable during restructuring phases.
- Track organizational changes at central AI vendors over time, not just as individual news items: a third dissolved safety team in two years is a pattern, not an isolated case, and patterns say more about future behavior than a single press release.
- Keep in mind the proximity between pre-IPO restructuring phases and safety cutbacks: economic pressure ahead of an IPO and building robust but costly internal control functions tend to pull in opposite directions - a dynamic that could repeat at other vendors in the same market.
The real value of this analysis isn't a verdict on whether OpenAI's current capabilities are inadequately reviewed - that can't be conclusively judged from the outside. It's making visible a pattern that connects three consecutive team dissolutions, two prominent departures, and an independent industry report: relying strategically on an AI vendor means watching its internal governance structure as closely as its model roadmap.