Researchers have introduced LCGuard, a latent communication guard designed to ensure safe key-value sharing in multi-agent systems that rely on large language models. This development is crucial as transformer-based models increasingly use intermediate communication to coordinate complex tasks, with latent communication through key-value caches showing promise in improving efficiency and preserving task-relevant information. However, these caches also encode sensitive information, posing significant security risks. LCGuard aims to mitigate these risks by safeguarding the sharing of key-value pairs, thereby preventing potential attacks that could compromise the entire system1. The security implications of large language model developments, particularly those leveraging transformer architectures, are significant and often overlooked in the hype surrounding their capabilities. As these models become more prevalent, the need for robust security measures like LCGuard will only continue to grow, making it essential for practitioners to prioritize the security of their multi-agent systems.