Reproducibility Package for "Attack Success and Utility Cost of Defenses Against Indirect Prompt Injection in Tool-Using Large Language Model Agents: A Controlled Empirical Evaluation"

This record contains the reproducibility package for the study “Attack Success and Utility Cost of Defenses Against Indirect Prompt Injection in Tool-Using Large Language Model Agents: A Controlled Empirical Evaluation.” It includes the experiment code, frozen result files, analysis materials, workbook, and supporting documentation used for the study. The study evaluates indirect prompt injection against tool-using LLM agents on the AgentDojo banking suite, comparing three current models under no-defense and defense conditions while measuring attack success rate, benign utility, and utility-under-attack. The package is intended to support verification of the reported results and manuscript preparation. Included materials comprise the manuscript draft, project handoff notes, experiment playbook, data workbook, and the main experiment bundle referenced in the project documentation. The handoff file identifies the canonical frozen runs and analysis outputs as the basis for the paper’s reported findings.