A significant breakthrough in large language models (LLMs) has been claimed by AI startup Subquadratic, which asserts it has overcome a longstanding mathematical bottleneck that has hindered LLM development for nearly a decade. By reducing the number of computations required for transformers to generate answers, Subquadratic's innovation promises to yield faster and more cost-effective LLMs. This development has substantial implications for the capability and risk profiles of LLMs, which are built on transformer architectures1. The security implications of such advancements are considerable, as they can introduce new vulnerabilities and expand existing attack surfaces. As LLMs continue to evolve, their potential risks and benefits will be reshaped, underscoring the need for ongoing evaluation and mitigation of potential security threats. This matters to practitioners because the trajectory of LLM development will significantly influence the security landscape of AI systems.