Adaptive Interleaved Reasoning (AIR) integrates code into multimodal large language models (MLLMs) to enhance their capabilities, building on the foundation laid by OpenAI's o3 model1. This approach deviates from the conventional focus on vision-perception tasks and predefined heuristics for visual manipulation. By incorporating code, AIR enables MLLMs to address numerical and symbolic reasoning tasks more effectively. The development of AIR is significant, particularly in the context of state-aligned activities involving OpenAI, as it shifts the threat model from criminal to geopolitical. This change necessitates a different strategy for mitigating potential risks. The introduction of AIR has the potential to elevate the capabilities of MLLMs, making them more versatile and powerful tools. As a result, practitioners must reassess their understanding of the threat landscape and adapt their approaches to address the emerging challenges posed by advanced language models, so what matters most is the ability to develop effective countermeasures against the potential misuse of these models.