As AI code assistants become increasingly integrated into software development workflows, understanding how their code compares to human-written programs is critical for ensuring reliability, maintainability, and security. In this paper, we present a large-scale comparison of code authored by human developers and three state-of-the-art LLMs, i.e., ChatGPT, DeepSeek-Coder, and Qwen-Coder, on multiple dimensions of software quality: code defects, security vulnerabilities, and structural complexity. Our evaluation spans over 500k code samples in two widely used languages, Python and Java, classifying defects via Orthogonal Defect Classification and security vulnerabilities using the Common Weakness Enumeration. We find that AI-generated code is generally simpler and more repetitive, yet more prone to unused constructs and hardcoded debugging, while humanwritten code exhibits greater structural complexity and a higher concentration of maintainability issues. Notably, AI-generated code also contains more high-risk security vulnerabilities. These findings highlight the distinct defect profiles of AI-and humanauthored code and underscore the need for specialized quality assurance practices in AI-assisted programming.

Deep Learning (DL)-based code generators have seen significant advancements in recent years. Tools such as GitHub Copilot are used by thousands of developers with the main promise of a boost in productivity. However, researchers have recently questioned their impact on code quality showing, for example, that code generated by DL-based tools may be affected by security vulnerabilities. Since DL models are trained on large code corpora, one may conjecture that low-quality code they output is the result of low-quality code they have seen during training. However, there is very little empirical evidence documenting this phenomenon. Indeed, most of previous work look at the frequency with which commercial code generators (e.g., Copilot, ChatGPT) recommend low-quality code without the possibility of relating this to their (publicly unavailable) training set. In this paper, we investigate the extent to which low-quality code instances seen during training affect the quality of the code generated at inference time. We start by fine-tuning a pre-trained DL model on a large-scale dataset ($>4.4 \mathrm{M}$functions) being representative of those usually adopted in the training of code generators. We show that 4.98 % of functions in this dataset exhibit one or more quality issues related to security, maintainability, coding practices, etc. We use the fine-tuned model to generate 551 k Python functions, showing that 5.85 % of them are affected by at least one quality issue. We then remove from the training set the low-quality functions, and use the cleaned dataset to fine-tune a second model which has been used to generate the same 551k Python functions. We show that the model trained on the cleaned dataset exhibits similar performance in terms of functional correctness as compared to the original model (i.e., the one trained on the whole dataset) while, however, generating a statistically significant lower number of low-quality functions (2.16 %). Our study empirically documents the importance of high-quality training data for code generators.

This practical experience report explores Neural Machine Translation (NMT) models’ capability to generate offensive security code from natural language (NL) descriptions, highlighting the significance of contextual understanding and its impact on model performance. Our study employs a dataset comprising real shellcodes to evaluate the models across various scenarios, including missing information, necessary context, and unnecessary context. The experiments are designed to assess the models’ resilience against incomplete descriptions, their proficiency in leveraging context for enhanced accuracy, and their ability to discern irrelevant information. The findings reveal that the introduction of contextual data significantly improves performance. However, the benefits of additional context diminish beyond a certain point, indicating an optimal level of contextual information for model training. Moreover, the models demonstrate an ability to filter out unnecessary context, maintaining high levels of accuracy in the generation of offensive security code. This study paves the way for future research on optimizing context use in AI-driven code generation, particularly for applications requiring a high degree of technical precision such as the generation of offensive code.

AI-based code generators have become pivotal in assisting developers in writing software starting from natural language (NL). However, they are trained on large amounts of data, often collected from unsanitized online sources (e.g., GitHub, HuggingFace). As a consequence, AI models become an easy target for data poisoning, i.e., an attack that injects malicious samples into the training data to generate vulnerable code.

Evaluating the correctness of code generated by AI is a challenging open problem. In this paper, we propose a fully automated method, named ACCA, to evaluate the correctness of AI-generated code for security purposes. The method uses symbolic execution to assess whether the AI-generated code behaves as a reference implementation. We use ACCA to assess four state-of-the-art models trained to generate security-oriented assembly code and compare the results of the evaluation with different baseline solutions, including output similarity metrics, widely used in the field, and the well-known ChatGPT, the AI-powered language model developed by OpenAI. Our experiments show that our method outperforms the baseline solutions and assesses the correctness of the AI-generated code similar to the human-based evaluation, which is considered the ground truth for the assessment in the field. Moreover, ACCA has a very strong correlation with the human evaluation (Pearson’s correlation coefficient r=0.84 on average). Finally, since it is a full y automated solution that does not require any human intervention, the proposed method performs the assessment of every code snippet in ∼0.17 s on average, which is definitely lower than the average time required by human analysts to manually inspect the code, based on our experience.

AI-based code generators are an emerging solution for automatically writing programs starting from descriptions in natural language, by using deep neural networks (Neural Machine Translation, NMT). In particular, code generators have been used for ethical hacking and offensive security testing by generating proof-of-concept attacks. Unfortunately, the evaluation of code generators still faces several issues. The current practice uses output similarity metrics, i.e., automatic metrics that compute the textual similarity of generated code with ground-truth references. However, it is not clear what metric to use, and which metric is most suitable for specific contexts. This work analyzes a large set of output similarity metrics on offensive code generators. We apply the metrics on two state-of-the-art NMT models using two datasets containing offensive assembly and Python code with their descriptions in the English language. We compare the estimates from the automatic metrics with human evaluation and provide practical insights into their strengths and limitations.