Prompt Injection: A Security Challenge for Modern AI Communication

Prompt injection attacks pose a multifaceted security threat to large language models (LLMs), undermining not just output content but also meaning, intentionality and trust in AI-mediated communication. Through qualitative analysis of documented text-based and multimodal attacks, this study reveals consistent patterns of tonal shift, narrative distortion and epistemic asymmetry that serve manipulative agendas. Drawing on concepts from data colonialism and surveillance capitalism, we argue these attacks constitute symbolic and ideological exploits rather than mere technical flaws. A layered defense model is proposed combining secure-by-design engineering, regulatory oversight and critical discourse analysis grounded in linguistics, media studies, and ethics, in order to detect and counter manipulative meaning constructions. An interdisciplinary security framework is essential to preserve the integrity, autonomy and societal trust in AI communication systems.