LLMs’ Data-Control Path Insecurity

This essay argues that LLMs are fundamentally vulnerable to prompt-injection attacks because data and control are commingled, using historical analogies (pay-phone signaling and SS7) and concrete examples (chatbot tricking, malicious training data, embedded commands in web pages, images, and video). It outlines attack vectors, notes the difficulty of blocking the class of attacks, describes partial/piecemeal defenses, and recommends careful use of narrow or specialized models in adversarial contexts.