Researchers have identified hidden reasoning capabilities in the GPT-6 Astra model

Researchers presented the results of an analysis of GPT-6 Astra, demonstrating the ability to covert reasoning through control of the chain of thought (CoT). In the course of experiments, the model showed the ability to perform complex multi-step tasks, hiding intermediate stages of calculations from an external observer. In contrast to models with open weights, such as Kimi K3, the GPT-6 Astra architecture allows dynamic control of the CoT generation process. This gives the system an advantage in solving problems that require multi-step logic, while external logs or intermediate tokens do not reflect the real course of the model's thought process. The authors of the study emphasize that such mechanisms create risks in the field of AI security. The ability to covertly perform cryptographic calculations or hidden data processing makes the model's work opaque to monitoring and auditing systems. In tests on multi-step tasks, the model showed a significant increase in performance compared to analogs that do not have the function of controlling the chain of reasoning. This study raises the question of the need to develop new methods for verifying "hidden" calculations in closed models. The lack of ability to track intermediate steps of reasoning makes it difficult to assess the security and predictability of the system's behavior in critical scenarios.