Abstract: The use of Large Language Models (LLMs) in augmented democracy systems is gaining traction in recent literature, as they help generate consensus statements and aggregate preferences from diverse citizen inputs. However, their integration introduces critical vulnerabilities that could compromise their reliability in democratic processes. This paper examines their robustness against prompt-injection attacks, where adversarial manipulations distort LLM-driven consensus formation by injecting malicious instructions in external inputs. Our preliminary findings indicate that SeqAlign, as a defense technique, can effectively mitigate these attacks, but further research is needed as new attack categories and strategies continue to emerge. These results highlight the need for strong safeguards to ensure the reliability and resilience of LLM-based augmented democracy systems.