This repository contains the code related to the paper DeVAIC: A Tool for Security Assessment of AI-generated Code accepted for publication in Information and Software Technology (IST) journal.

Description

DeVAIC (Detection of Vulnerabilities in AI-generated Code) is a fast static analysis tool for detecting vulnerabilities in code written in Python language. It can work even on code snippets, i.e. incomplete code due to the lack of initial import statements, single function definition, etc. It detects vulnerabilities belonging to the OWASP categories listed in the Top 10 of 2021 (i.e., Broken Access Control, Cryptographic Failures, Injection, Insecure Design, Security Misconfiguration, Vulnerable and Outdated Components, Identification and Authentication Failures, Software and Data Integrity Failures, Security Logging and Monitoring Failures, and SSRF).

🛠️ Step 1: Initial Setup

🚨 Prerequisites:

• Please run on a Linux OS or macOS. For Windows users, you can utilize the Windows Subsystem for Linux (WSL); in this case, please ensure to have the WSL installed before proceeding.
• You must have Python 3.8 or later installed on the environment where you launch DeVAIC.

Make the file executable with the following commands:

chmod +x devaic.sh

chmod +x tool_derem.sh

For macOS users:

In the case of macOS, type the following command from the shell to use the GNU-like version of grep by ensuring compatibility with the grep command:

brew install grep

🚀 Step 2: Run the experiments

Input file

Move the file to analyze (e.g., YOUR_INPUT_FILE.txt) into the directory DeVAIC/input.

⚠️ Disclaimer

WARNING: Each code snippet in the input file must be written line by line. It is recommended to use the YOUR_INPUT_FILE in .txt format.

For instance, the input folder contains four files in txt format each having the code snippets generated by four different models, i.e., GitHub Copilot (github_copilot.txt), Google Gemini (google_gemini.txt), Microsoft Copilot (microsoft_copilot.txt) and OpenAI ChatGPT (openai_chatgpt.txt).

Running DeVAIC

To launch the detection tool, move into the main folder and run the following command:

./devaic.sh input/[YOUR_INPUT_FILE.txt]

At the end of execution, the tool generates a report file which can be found at path DeVAIC/results/detection/DET_[timestamp]_[YOUR_INPUT_FILE].txt. This report contains information for each examined snippet as follows:

1. If the snippet is evaluated as vulnerable, the following information will be provided:

   • A label "(!) VULN CODE" indicating that one or more vulnerabilities were detected in the snippet.
   • The execution time taken by the rules on the single snippet.
   • The list of OWASP categories associated with the vulnerabilities detected in the snippet.
   • Finally, the snippet itself.

2. If no vulnerabilities are detected in the snippet, the following information will be reported:

   • A label "==> SAFE CODE".
   • The execution time taken by the rules on the single snippet.
   • Finally, the snippet itself.

Interpreting Results

At the end of its execution, in addition to the creation of the DET file described above, DeVAIC displays the following information in the Command Prompt from which it was launched:

Label on prompt	Meaning
#DimTestSet	Total number of evaluated snippets
#TotalVulnerabilities	Number of vulnerable snippets detected
#SafeCode	Number of snippets marked as safe
Vulnerability Rate	Rate of detected vulnerabilities (i.e. number of vulnerable snippets out of total snippets)
List of OWASP categories	Number of vulnerable snippets belonging to each OWASP category
Runtime	Overall execution time on the entire dataset of snippets
Average runtime per snippet	Average execution time per single snippet

💻 Practical Usage Example

1. To detect the vulnerabilities among the snippets listed in github_copilot.txt located in the input folder, move into the main folder and use the following command:

./devaic.sh input/github_copilot.txt

2. Then, move to the path DeVAIC/results/detection to analyze the results of the detection shown in the file DET_[timestamp]_github_copilot.txt.

📊 Manual Analysis Results: The utils folder contains an Excel file with the manual analysis of the code samples from the input folder. The Excel file has four sheets, each containing the 125 code samples generated by each of the four models (i.e., GitHub Copilot, Google Gemini, Microsoft Copilot, and OpenAI ChatGPT). Each row in a sheet contains the manual analysis of the corresponding line in the respective text file in the input folder.

Citation

If you use DeVAIC in academic context, please cite it as follows:

@article{COTRONEO2025107572,
title = {DeVAIC: A tool for security assessment of AI-generated code},
journal = {Information and Software Technology},
volume = {177},
pages = {107572},
year = {2025},
issn = {0950-5849},
doi = {https://doi.org/10.1016/j.infsof.2024.107572},
url = {https://www.sciencedirect.com/science/article/pii/S0950584924001770},
author = {Domenico Cotroneo and Roberta {De Luca} and Pietro Liguori},
keywords = {Static code analysis, Vulnerability detection, AI-code generators, Python}
}