Dataset for "Implementation and Evaluation of OWASP LLM Security Mitigations in Multi-Agent Systems: A Design Science Approach"

This dataset constitutes the primary software and evaluation artifacts for the MSc thesis "Implementation and Evaluation of OWASP LLM Security Mitigations in Multi-Agent Systems: A Design Science Approach"

The Excel file in this Zenodo record captures the quantitative evaluation results (including WildJailbreak) that underpin the empirical findings reported in Chapters 3 and 5 of the thesis.

The Excel file provides:

Per-scenario results for internal OWASP-aligned security tests (LLM01 prompt injection, LLM02 sensitive data protection, LLM06 excessive agency, LLM09 misinformation), including scenario descriptions, categories, severity, detector decisions, and outcome labels (e.g., “Correctly Blocked”, “Correctly Allowed”).

Aggregated performance tables (overall metrics and per-category metrics) with detection counts, false positives, precision, recall, F1 scores, and 95% confidence intervals, including separate rows for the WildJailbreak adversarial benchmark (2,210 prompts).

Baseline vs FPAS comparison summaries, ablation study configurations (e.g., with/without WildJailbreak-informed patterns and opinion detection), and an error taxonomy describing typical failure modes (obfuscation, indirect attacks, context dilution, multilingual attacks).

McNemar test results quantifying the statistical significance of the improvement from the baseline pattern detector to the enhanced FPAS configuration, plus metadata about file generation time and totals.