OpenAI Quietly Dissolved Its Team That Watched for AI Going Rogue
The group whose whole job was to spot dangers in OpenAI's own models is gone. Safety critics say the timing is telling.
Key points
- OpenAI disbanded its dedicated preparedness team at the end of April 2025, according to the Financial Times.
- The preparedness team tested whether AI models posed serious risks, such as helping build weapons or hacking other systems unpredictably.
- Responsibility for those risk checks has been split by topic, biology, cybersecurity, and folded into teams that already existed.
- Several senior safety figures have left OpenAI recently, including ethics lead Chloé Bakalar and head of safety Johannes Heidecke.
- The team's former head, Dylan Scandinaro, now focuses on so-called "recursive self-improving" AI, meaning systems that could rewrite and upgrade themselves without human input.
OpenAI no longer has a single team whose purpose was to find out whether its own models could turn dangerous.
The preparedness team, as it was known inside the company, was responsible for red-teaming OpenAI's models. Red-teaming means deliberately trying to make a system misbehave, the way a security researcher tries to break into a computer before a real attacker does. The team looked at worst-case scenarios: could a model help someone build a biological weapon? Could it hack another company on its own? When the answers looked alarming, the team was supposed to develop ways to pull the risk back.
Now that team is gone.
First reported by the Financial Times, the decision came at the end of last month. OpenAI didn't eliminate the work entirely. Instead, it broke the responsibilities into specific subject areas and handed each piece to existing departments. No single group holds the whole picture anymore.
Why does this matter to ordinary people?
Powerful AI models are already embedded in tools people use at work, in search engines, in customer-service bots. The preparedness team was one of the internal guardrails designed to catch problems before those tools reach the public. Fewer dedicated safety staff means the checks happen differently. Not necessarily less, but the structure that existed specifically to sound an alarm is gone.
Jan Leike, a prominent AI safety researcher who resigned from OpenAI in 2024, told the Financial Times the company was pushing "shiny products" ahead of safety considerations. The preparedness team's dissolution is the latest structural change in a longer pattern. OpenAI has already wound down its AGI Readiness team (AGI stands for artificial general intelligence, meaning a system that could match or surpass human reasoning across many tasks) and its Superalignment team, which studied how to keep very powerful future AI systems under human control. Bakalar, Heidecke, and Chief Futurist Josh Achiam have all departed in recent months.
What happens next?
Dylan Scandinaro, who joined OpenAI from rival lab Anthropic in February, will now concentrate on recursive self-improving AI: systems capable of rewriting their own code to become more capable without human instruction. That's arguably the highest-stakes problem in the field right now. We covered Lilian Weng's return to OpenAI to lead exactly that kind of research on 29 July 2026, and the concentration of talent on self-improving systems, inside OpenAI and across the industry, is accelerating faster than the safety structures around it.
OpenAI is heading toward what analysts widely expect to be a large IPO, a stock-market listing that would let outside investors buy shares. Whether pressure to impress future investors is pushing safety work to the margins is a fair question, and OpenAI hasn't publicly commented on the specifics of the reorganisation.
