PROMPT INJECTION ATTACKS IN LARGE LANGUAGE MODELS VIA A COMPREHENSIVE ANALYSIS OF ATTACK VECTORS, DEFENSE MECHANISMS, AND FUTURE DIRECTIONS

We present an extended taxonomy of 7 attack vectors and formalize prompt injection as an adversarial optimization problem.Empirical evaluation on GPT-4 (0613), Claude 3 Opus, and Gemini 1.5 Pro using a curated dataset of 1,000 validated attack scenarios shows baseline safeguards provide ~58% ± 3.2% protection (42% ASR).State-ofthe-art single-layer defenses reach at most ~82% but with elevated false positives.Our main contribution is an integrated six-layer defense-in-depth framework (building on but extending prior techniques like spotlighting [15] and multi-model verification [16]), which reduces ASR to 3.2% ± 1.1% (p < 0.001 vs. baseline) with acceptable false-positive rates (~5.3%) and moderate latency.Ablation confirms architectural separation and multimodel verification as key contributors.We analyze five real-world incidents (2023-2025) and provide deployment guidelines.