Raphaël Khoury

LLM models are increasingly used to generate code, but the quality and security of this code are often uncertain. Several recent studies have raised alarm bells, indicating that such AI-generated code may be particularly vulnerable to cyberattacks. However, most of these studies rely on code that is generated specifically for the study, which raises questions about the realism of such experiments. In this study, we perform a large-scale empirical analysis of real-life code generated by ChatGPT. We evaluate code generated by ChatGPT both with respect to correctness and security and delve into the intentions of users who request code from the model. Our research confirms previous studies that used synthetic queries and yielded evidence that LLM-generated code is often inadequate with respect to security. We also find that users exhibit little curiosity about the security features of the code they ask LLMs to generate, as evidenced by their lack of queries on this topic.

ChatGPT, a Large Language Model (LLM) maintained by OpenAI, has demonstrated a remarkable ability to seemingly comprehend and contextually generate text. Among its myriad applications, its capability to autonomously generate and analyze computer code stands out as particularly promising. This functionality has piqued substantial interest due to its potential to streamline the software development process. However, this technological advancement also brings to the forefront significant apprehensions concerning the security of code produced by LLMs. In this article, we survey recent research that examines the use of ChatGPT to generate secure code, detect vulnerabilities in code, or perform other tasks related to secure code development. Beyond categorizing and synthesizing these studies, we identify important insights into ChatGPT’s potential impact on secure programming. Key findings indicate that while ChatGPT shows great promise as an aid in writing secure code, challenges remain. Its effectiveness varies across security tasks, depending on the context of experimentation (programming language, CWE, code length, etc.) and the benchmark used for comparison–whether against other LLMs, traditional analysis tools, or its own versions. The overall trend indicates that GPT-4 consistently surpasses its predecessor in most tasks.

In recent years, there has been a steady increase in the number of reported vulnerabilities (CVEs), increasing the workload of organizations trying to update their systems promptly. This underscores the need to prioritize specific critical vulnerabilities over others to effectively prevent cyberattacks. Unfortunately, the current methods available for assessing the exploitability of vulnerabilities have substantial shortcomings. In particular, they often consist in prediction models that encode data that may not be immediately available at the time a vulnerability is first reported. In this paper, we introduce an innovative exploitability prediction method that exclusively uses information accessible at the time of a CVE’s initial publication. Our approach outperforms the most widely used vulnerability exploit prediction algorithms in scenarios where data is subject to the aforementioned limitations.

The detection of malicious Android applications is a major security challenge. A number of machine learning-based techniques have been put forth, and some of them have attained great accuracy. However, the diversity of apps and frequency at which new malware families are found means that the issue remains unresolved. In this paper, we use both static, dynamic and hybrid analysis to automatically classify Android apps as benign or infected. We compare all three approaches on a common dataset — the TwinDroid dataset which contains over 15,000 system call traces from over 9,000 benign and infected app. This method allows comparison on equal footing. We make further contributions on the topic of feature selection and trace abstraction.

Over the past several years, several governments have, at times, pushed for the idea that commercial software should be required to include a “backdoor,” a deliberate vulnerability whose existence and exploitation mechanism are disclosed only to the appropriate authorities. This would enable the authorities to obtain access to the information contained in any device running this software when needed to react to criminal activity.