OpenAI assigned the Astra model a Critical level of cybersecurity

Company OpenAI officially introduced the GPT-6 Astra model, which became the first product of the organization to receive the Critical level within the internal Preparedness Framework evaluation system. This classification means that the neural network has the ability to independently find unknown vulnerabilities and develop scenarios for their exploitation.

According to the report, the model is capable of performing tasks to find gaps in defense without constant human control, requiring only the setting of a basic goal. During the development of GPT-6 Astra, engineers were forced to intentionally slow down the training stages to implement additional control mechanisms and prevent potential misuse of the system's capabilities.

The assignment of the critical security level was a consequence of the risk assessment associated with autonomous search for attack vectors. The documentation states that current protection protocols do not always keep up with the pace of model functionality growth, which forces the company to revise its approaches to the release cycle and coordination with other laboratories.

Against the backdrop of these events, the management of leading AI companies began to publicly discuss the need to slow down the pace of the arms race in the field of AI. In particular, industry representatives point to the gap between the speed of model development and the development of security systems, which has already led to adjustments to IPO plans and changes in market expectations in the artificial intelligence technology sector.