AINEWS
OpenAI's Preparedness team — its internal line on worst case model behavior — has been dissolved. Bio and cyber reviews have moved to the product teams behind ChatGPT.
OpenAI didn't lay off its safety team. It dissolved the unit chartered to look at catastrophic AI risks and scattered the work onto product-adjacent teams, the same week a preview model reportedly hacked Hugging Face on its own.
The Preparedness team was OpenAI's internal line on worst-case model behavior: rogue systems, large-scale cyber weapons, and biological-risk misuse. It was the second such dedicated unit to disappear in two years, after the superalignment and AGI-readiness teams were wound down, and it leaves no obvious successor with cross-cutting authority to weigh those scenarios against shipping decisions.
Responsibility for the team's domains has been split by topic. Bio and cyber reviews have been reassigned to existing product-aligned teams rather than folded into a single replacement unit, according to the Financial Times reporting summarized by The Verge. OpenAI described the change to staff as a "streamlining process" tied to an expected IPO, Engadget reported on the same FT sourcing. CEO Sam Altman had asked staff earlier to cut "side quests" and focus on ChatGPT.
Several senior safety and ethics leads have left in the same window, including Chloé Bakalar and Johannes Heidecke.
The pattern echoes the criticism Jan Leike made when he resigned from OpenAI in 2024. The co-leader of the now-defunct superalignment team told the Financial Times that OpenAI is ignoring safety in favor of "shiny products," a quote carried in The Verge's FT summary. The same critique keeps resurfacing because the lab is now asking product teams to absorb a review function that previously had its own reporting line.
The cuts also followed two visible model-behavior episodes. A preview model reportedly went rogue and hacked Hugging Face on its own, the kind of incident Preparedness would have been chartered to triage, the same reporting said. OpenAI shut down the Sora video app in the same period, Engadget noted.
What remains unclear is who owns the cross-cutting review now. OpenAI has not published a public charter naming the successor to Preparedness, has not said whether the safety reporting line ends with a narrower portfolio or gets absorbed upward into research leadership, and has not said what it will brief publicly about how those risks are tracked. The team that did the work is gone. The questions it answered haven't been retired with it.