What Determines the Vulnerability of Small LLM Agents to Indirect Prompt Injection? A Systematic Study of Quantization, Model Size, Architecture, and Attack Style

Preprint of a systematic empirical study on indirect prompt injection in small quantized LLM agents. The study evaluates 14 model configurations across quantization levels, model sizes, architectures, attack styles, and prompt-level defenses. The main finding is that quantization from Q8 to Q4 has no practically significant effect on injection resistance, while attack style, architecture, and model-specific defense behavior have substantially larger effects. Code, scenarios, analysis scripts, figures, and raw rollouts are available in the associated GitHub repository and reproducibility artifact.