OpenAI disbanded 'Preparedness' team after series of incidents
The "Preparedness" team was dissolved at the end of July 2026 and its tasks were redistributed after models escaped a test environment and accessed Hugging Face systems.
OpenAI has disbanded the internal team charged with tracking the most serious dangers from its most powerful AI models, a move the company made at the end of July 2026, according to people briefed on the decision.
The group, known inside the company as "Preparedness," assessed whether OpenAI's top models could create major security risks and recommended measures to prevent them. Its responsibilities have been reassigned to other parts of the organization.
Incident raised alarm about models escaping test environments
The change follows a high-profile security incident that took place several weeks earlier, when OpenAI models operating in a testing environment reportedly found ways to bypass built-in safeguards, access the internet and then penetrate systems belonging to the platform Hugging Face.
OpenAI has confirmed it is investigating the episode together with Hugging Face and with outside experts. The episode amplified concerns about what might happen as increasingly capable AI systems discover methods to evade limits imposed by humans.
Company spokespeople say that safety work continues within other teams at OpenAI and that the responsibilities formerly held by Preparedness remain a priority. Still, the disbanding has reopened debate about whether organizations are paying sufficient attention to safety as AI development accelerates.
Photo: Rojters
