PromptAudit: Systematically Evaluating the Impact of Prompt Engineering on Code Vulnerability Detection

In the field of AI security, accurately evaluating large language models' (LLM) ability to detect code vulnerabilities has always been a core challenge. PromptAudit is an end-to-end experimental platform specifically designed for systematically studying the impact of prompt engineering techniques on code security classification. By fixing variables such as datasets and model backends and only changing prompt strategies, it enables controlled comparative experiments, helping researchers understand the real impact of prompt strategies on vulnerability detection performance.

With the widespread application of LLMs in code analysis and security auditing, the same model shows significant differences in vulnerability detection accuracy under different prompt strategies. However, the industry lacks standardized tools to isolate these differences. PromptAudit fills this gap by fixing datasets, model backends, decoding configurations, and reporting processes, and only changing prompt strategies, output protocols, and parsing modes to achieve controlled comparative experiments.

PromptAudit supports various prompt ablation experiments:

• Zero-shot: Direct classification without providing examples
• Few-shot: Classification with a small number of example cases
• CoT: Guiding the model to answer through reasoning
• Adaptive CoT: More guided reasoning prompts
• Self-consistency: Majority voting from multiple samples
• Self-verification: Reasoning → Verification → Conclusion

Additionally, it supports ablation tests for output protocols (verdict_first/last) and parsing modes (strict/structured/full).

PromptAudit adopts a modular design:

• Dataset Layer: Supports Hugging Face, local CVE, and toy datasets
• Model Layer: Compatible with API models, Hugging Face local models, and Ollama services
• Prompt Layer: Plug-and-play prompt strategies for easy expansion
• Evaluation Layer: Label parsing, metric calculation, and report generation
• UI Layer: Tkinter graphical interface for experiment monitoring

Experiments generate timestamped artifact directories (metrics.csv, report.html, etc.) to ensure results are traceable and reproducible.

The platform provides comprehensive operation control:

• Pause/Resume: Pause after completing the current sample and save checkpoints
• Checkpoint Resume: Restore from the latest checkpoint on disk
• Safe Stop: Stop at boundaries and generate partial artifacts
• Anti-sleep Mode: Prevent the system from sleeping during experiments

These features support resource management for experiment cycles ranging from hours to days.

Limitations of PromptAudit:

• CVE-related datasets have label noise derived from patches
• Vulnerability judgment of code snippets lacks runtime context
• Migration issues of small open-source model results to proprietary systems

It is recommended that papers discuss these limitations and address them through additional experiments or strict subset selection.

Smoke test process: Select the mistral:latest model, zero_shot prompt, toy dataset, verdict_first protocol, and full parsing mode to generate a report in a few minutes.

Application scenarios:

• Academic research: Compare the performance of prompt strategies in security classification
• Industrial applications: Evaluate the improvement of prompt schemes on internal audit tools
• Teaching demonstrations: Show the impact of prompt engineering on LLM outputs

PromptAudit provides a professional, controllable, and reproducible experimental platform for the LLM security community. It isolates prompt engineering variables, helps researchers accurately understand the impact of prompt strategies on vulnerability detection performance, and promotes the development of safer and more reliable AI-driven code audit tools.