A Preliminary Study on Using Large Language Models in Software Pentesting
(2401.17459)Abstract
Large language models (LLM) are perceived to offer promising potentials for automating security tasks, such as those found in security operation centers (SOCs). As a first step towards evaluating this perceived potential, we investigate the use of LLMs in software pentesting, where the main task is to automatically identify software security vulnerabilities in source code. We hypothesize that an LLM-based AI agent can be improved over time for a specific security task as human operators interact with it. Such improvement can be made, as a first step, by engineering prompts fed to the LLM based on the responses produced, to include relevant contexts and structures so that the model provides more accurate results. Such engineering efforts become sustainable if the prompts that are engineered to produce better results on current tasks, also produce better results on future unknown tasks. To examine this hypothesis, we utilize the OWASP Benchmark Project 1.2 which contains 2,740 hand-crafted source code test cases containing various types of vulnerabilities. We divide the test cases into training and testing data, where we engineer the prompts based on the training data (only), and evaluate the final system on the testing data. We compare the AI agent's performance on the testing data against the performance of the agent without the prompt engineering. We also compare the AI agent's results against those from SonarQube, a widely used static code analyzer for security testing. We built and tested multiple versions of the AI agent using different off-the-shelf LLMs -- Google's Gemini-pro, as well as OpenAI's GPT-3.5-Turbo and GPT-4-Turbo (with both chat completion and assistant APIs). The results show that using LLMs is a viable approach to build an AI agent for software pentesting that can improve through repeated use and prompt engineering.
Overview
-
The study investigates the potential of LLMs in identifying software security vulnerabilities through prompt engineering.
-
Researchers utilized the OWASP Benchmark Project 1.2, training AI agents built on various LLMs against SonarQube, a static application security testing tool.
-
Findings indicate that LLM-based AI agents can improve vulnerability identification accuracy with tailored prompt engineering.
-
The study suggests a promising path for integrating LLMs into software pentesting, emphasizing the importance of iterative prompt refinement.
Assessing LLMs' Efficacy in Software Pentesting Through Prompt Engineering
Introduction
The realm of cybersecurity has increasingly turned towards the incorporation of artificial intelligence to bolster defenses against pervasive threats. Particularly in software pentesting, where the identification of vulnerabilities in code is paramount, the potential of LLMs for automating security tasks has been a subject of considerable interest. In responding to this emerging frontier, a study conducted by researchers at the University of South Florida delved into evaluating the capability of LLMs in identifying software security vulnerabilities, under the hypothesis that LLM-based AI agents could progressively improve their performance in specific security tasks through interactive prompt engineering with human operators.
Background and Motivation
Software pentesting forms an essential part of secure software development practices, aimed at uncovering vulnerabilities before they can be exploited maliciously. Traditional tools used in pentesting, such as static application security testing (SAST) tools, often return a high volume of false positives, leading to pentester fatigue. Conversely, LLMs hold promise due to their capability for nuanced reasoning and potential for on-the-fly training through prompt adjustments. This study is motivated by the anticipation that LLMs, through iterative refinements in the prompts they are fed, could adapt to pentesting tasks with increasing acumen, providing a sustainable, dynamic approach to software security analysis.
Methodology
The study utilized the OWASP Benchmark Project 1.2, featuring 2,740 source code test cases with known vulnerabilities, to train and evaluate several AI agents built upon different off-the-shelf LLMs, including Google's Gemini-pro, OpenAI's GPT-3.5-Turbo, and GPT-4-Turbo. The performance of these AI agents was measured against that of SonarQube, a widely used SAST tool, both with and without the application of prompt engineering. The process involved the division of the benchmark's test cases into training and testing sets, with the engineered prompts being refined based on the AI agents' performance on the training set, then evaluated on the unseen testing set.
Results
The results from this exploratory study reveal a nuanced landscape:
- AI agents leveraging LLMs exhibited potential for improving their accuracy in identifying software vulnerabilities through prompt engineering.
- In particular, the GPT-4-Turbo model utilizing the Assistants API demonstrated notable improvements post-prompt engineering, performing on par with or better than SonarQube in a majority of vulnerability categories.
- The study identified two primary categories wherein LLMs tended to err under base prompts: misinterpretations related to code flow and the identification of weak algorithms. Through prompt engineering, these errors were significantly mitigated.
Implications and Future Directions
This investigation substantiates the viable trajectory for LLMs in automating aspects of software pentesting, with prompt engineering serving as a critical mechanism for adapting these models to the nuances of security analysis. The findings underscore the importance of tailored prompt engineering strategies, catering to the specific strengths and weaknesses of different LLMs. Looking ahead, the study paves the way for a more in-depth exploration into the bounds of LLM capabilities in security-centric applications, including the potential for real-time learning and adaptation in the face of evolving software development practices and emerging threat landscapes.
Conclusion
The preliminary study by the University of South Florida offers a promising outlook on the integration of LLMs into the domain of software pentesting, highlighting the role of prompt engineering in enhancing the accuracy and adaptability of AI agents. As cybersecurity moves increasingly towards automation, the insights garnered from this research could inform the development of more effective, dynamic tools for securing software in an ever-evolving digital age.
This research was partially supported by the National Science Foundation and the Office of Naval Research, reflecting a broad interest in advancing the frontiers of cybersecurity through the integration of cutting-edge artificial intelligence technologies.
Create an account to read this summary for free: