OpenAI Shut Down Its Third Safety Team in Two Years

OpenAI quietly disbanded its Preparedness team at the end of July 2026, eliminating the group responsible for evaluating its most dangerous failure scenarios before new models ship. It is the third dedicated safety unit the company has dissolved in roughly two years, following the AGI Readiness team in 2024 and Mission Alignment in February 2026. The timing is hard to ignore: the decision came just days after a Hugging Face hacking incident in which OpenAI models ended up outside their controlled environment.

What the team actually did

Preparedness was not a general ethics group. According to Financial Times, its specialists focused on extreme, hard-to-reverse risks: large-scale cyberattacks, biological threats, and scenarios where an AI system gains enough autonomy to act outside its creator's control. That narrow, high-stakes mandate made it distinct from other internal safety work.

With the team gone, its responsibilities have been split across existing groups. Dylan Scandinaro, who led parts of the team's work, is now focused specifically on recursive self-improving AI — systems that can rewrite and upgrade themselves. OpenAI says the Preparedness Framework, its published standard for evaluating frontier models, remains in place.

Distributed safety, concentrated risk

OpenAI's replacement structure relies on its Frontier Governance Framework, which covers chemical, biological, radiological, and nuclear (CBRN) threats alongside manipulation risks and AI autonomy concerns. The company argues that spreading safety responsibilities across the whole infrastructure is more robust than concentrating them in one team.

Critics read it differently. When a single group owns catastrophic-risk assessment, accountability is clear. Distribute that function across teams with competing operational priorities, and it becomes easier for worst-case questions to fall through the gaps — especially as Cryptonomist notes that industry watchers are tracking safety staffing as a signal ahead of OpenAI's expected IPO.

The practical concern is institutional knowledge. The ethics chief, chief futurist, and safety lead have all left the company in close succession. The people who built the original risk models, ran the scenario exercises, and understood where the framework's blind spots were are no longer in the building. Their replacements, however capable, are starting from a lower base.

OpenAI has not published a detailed explanation of how the new distributed model will handle pre-release risk assessments. Until it does, the gap between what the Preparedness Framework promises on paper and what any team can operationally deliver remains an open question.