Security and Risk Implications of Transformer-Based Large Language Models

The growing adoption of Large Language Models (LLMs) across a wide range of applications has introduced new challenges in application security, complementing traditional deterministic software approaches rather than replacing them entirely. While this transition enables advanced capabilities in automated reasoning and content generation, it also exposes architectural characteristics that remain insufficiently addressed by many existing cybersecurity frameworks.A core security challenge arises from the use of natural language as a unified representational medium for both instructions and data within modern generative systems. Unlike classical computing architectures, where a strict separation exists between executable instructions and passive data, LLMs process system instructions and user inputs as undifferentiated sequences of natural-language tokens. This architectural property introduces a form of semantic ambiguity in which untrusted data may be interpreted as instructions under certain contextual conditions, leading to unintended model behaviour.This paper examines the security and risk implications of this ambiguity through architectural, operational, and legal lenses. Particular attention is given to agent-based systems and Retrieval-Augmented Generation (RAG) pipelines, where indirect prompt injection, tool abuse, and excessive model agency expand the attack surface beyond purely digital environments into physical systems and complex organizational supply chains. The analysis highlights the limitations of existing security assumptions and motivates the need for systemic resilience strategies tailored to LLM-driven systems.