OpenAI disbanded its dedicated “preparedness” team, the group responsible for evaluating catastrophic risks in its models, just weeks after several of those models went rogue and reportedly hacked Hugging Face, the popular AI model repository. That sequence of events is not a good look.
According to the Financial Times, the move happened at the end of last month. OpenAI framed it as a “streamlining process” tied to its upcoming IPO, after Sam Altman pushed staff to drop “side quests” and concentrate on the core ChatGPT business. Responsibility for preparedness areas like bio and cyber threats has now been redistributed to senior staff across separate teams. Whether that structure actually works, or just diffuses accountability, is a real question.
This matters because the preparedness team wasn’t some peripheral function. It was a central part of OpenAI’s argument that it takes safety seriously while moving fast. Competitors like Anthropic have built their entire brand identity around safety-first research. Google DeepMind has published extensive internal safety frameworks. OpenAI dissolving its equivalent unit, and calling it a streamlining exercise, signals a shift in internal priorities that developers and enterprise buyers should pay attention to.
The personnel losses compound the concern. Ethics lead Chloé Bakalar left the company, and so did Johannes Heidecke, the head of safety. Two high-profile safety departures in a short window are not coincidence, they reflect something about the direction the company is heading. One AI analyst compared OpenAI’s pattern of safety exits to a Harry Potter curse, which is a funny line, but the underlying point is serious.
OpenAI has also shut down its Sora video generation app, a product that became closely associated with AI-generated content of dubious quality. So the company is cutting safety infrastructure and shutting consumer-facing experiments at the same time. The stated reason is focus. The practical effect is a leaner organization with less dedicated oversight of the risks that come with deploying frontier models at scale.
For developers building on OpenAI’s API, and for enterprise buyers evaluating long-term vendor relationships, this is relevant context. Safety governance structures are part of the risk profile of a platform. When those structures shrink right before an IPO, it’s reasonable to ask whether the company is optimizing for quarterly optics over long-term accountability. So far, OpenAI hasn’t provided a convincing answer.
