OpenAI disbanded its preparation team at the end of last month, which was responsible for assessing the serious security risks of the artificial intelligence models it developed. The team, operating under the name “Preparedness”, was identifying the risks that could be caused by the ever-developing capabilities of the models and working on how to reduce them. Among the scenarios examined, there were situations such as an artificial intelligence system going out of control and gaining capabilities that could infiltrate another company’s systems. It is not stated that the company has completely abandoned its security efforts; instead, responsibility for specific risk areas, such as biological threats and cybersecurity, is shared among existing teams. Thus, it becomes clear that the preparatory work, which was previously carried out under a single team, has been distributed to different units within the organization.
The change in question is a continuation of the restructuring of OpenAI’s teams working on security and long-term artificial intelligence risks over the past few years. The company previously disbanded the AGI Readiness team, which focused on preparations for artificial general intelligence, and the Superalignment team, which researched how advanced artificial intelligence systems remain compatible with human goals. In addition, there were notable differences in the company’s security and research staff. Chloé Bakalar, head of ethics studies, Josh Achiam, who holds the title of Chief Futurist, and Johannes Heidecke, head of security studies, are among the names that have recently left OpenAI. Financial Times points out that these changes coincide with a period when the company is moving towards its possible IPO, which is expected to be quite large, and is making changes to its corporate structure.
OpenAI distributes AI security efforts across different teams
Changes in OpenAI’s security structure also revived criticism from some researchers who had previously left the company. Jan Leike, one of the leaders of the Superalignment team, left OpenAI in 2024 and criticized the company’s security culture at that time. In his assessment to the Financial Times, Leike argued that OpenAI prioritizes developing “shiny products” over security. This view is a criticism of the company’s current security approach and does not have the same meaning as OpenAI’s official position. The fact that the duties of the preparation team are transferred to existing units in areas such as biological security and cyber security, rather than being completely abolished, shows that the company continues to address these issues with a different organizational model.
The role of Dylan Scandinaro, who is the head of the preparation team, also changes with this restructuring. Scandinaro, who moved to OpenAI from rival artificial intelligence company Anthropic in February, will now focus on the possible effects of “recursively self-improving” artificial intelligence systems. This concept expresses the possibility that an artificial intelligence system can contribute to the process of improving its own capabilities and reuse the higher capacity it has achieved for further developments. To what extent such a mechanism will be possible and at what speed it can be realized are among the controversial research topics in the field of artificial intelligence security. Scandinaro’s new workspace also reveals that OpenAI has not completely terminated its long-term risk research, but continues them under a different organizational structure.
This change to OpenAI’s security organization raises the question of how independent risk assessment will be conducted, especially as the capabilities of the company’s models increase. Putting risk areas directly under the responsibility of expert teams can benefit from deeper expertise on issues such as biological security or cyber security. On the other hand, the disappearance of a central structure such as a preparedness team may raise questions about the power of an in-house mechanism that evaluates different risk categories together. The determining factor here will be the authority, personnel and resources with which the distributed responsibilities will be maintained in the existing teams. The balance between the company’s product development pace and its efforts to measure the risks that advanced models may pose will continue to be one of the key criteria in evaluating OpenAI’s security approach.
Join Channel