Tamás Bisztray

Norbert Tihanyi, Tamas Bisztray, Ridhi Jain et al.

This paper presents the FormAI dataset, a large collection of 112, 000 AI-generated compilable and independent C programs with vulnerability classification. We introduce a dynamic zero-shot prompting technique constructed to spawn diverse programs utilizing Large Language Models (LLMs). The dataset is generated by GPT-3.5-turbo and comprises programs with varying levels of complexity. Some programs handle complicated tasks like network management, table games, or encryption, while others deal with simpler tasks like string manipulation. Every program is labeled with the vulnerabilities found within the source code, indicating the type, line number, and vulnerable function name. This is accomplished by employing a formal verification method using the Efficient SMT-based Bounded Model Checker (ESBMC), which uses model checking, abstract interpretation, constraint programming, and satisfiability modulo theories to reason over safety/security properties in programs. This approach definitively detects vulnerabilities and offers a formal model known as a counterexample, thus eliminating the possibility of generating false positive reports. We have associated the identified vulnerabilities with Common Weakness Enumeration (CWE) numbers. We make the source code available for the 112, 000 programs, accompanied by a separate file containing the vulnerabilities detected in each program, making the dataset ideal for training LLMs and machine learning algorithms. Our study unveiled that according to ESBMC, 51.24% of the programs generated by GPT-3.5 contained vulnerabilities, thereby presenting considerable risks to software safety and security.

Ryan Marinelli, Josef Pichlmeier, Tamas Bisztray

In this work, we propose a metric called Number of Thoughts (NofT) to determine the difficulty of tasks pre-prompting and support Large Language Models (LLMs) in production contexts. By setting thresholds based on the number of thoughts, this metric can discern the difficulty of prompts and support more effective prompt routing. A 2% decrease in latency is achieved when routing prompts from the MathInstruct dataset through quantized, distilled versions of Deepseek with 1.7 billion, 7 billion, and 14 billion parameters. Moreover, this metric can be used to detect adversarial prompts used in prompt injection attacks with high efficacy. The Number of Thoughts can inform a classifier that achieves 95% accuracy in adversarial prompt detection. Our experiments ad datasets used are available on our GitHub page: https://github.com/rymarinelli/Number_Of_Thoughts/tree/main.

Rebeka Tóth, Tamas Bisztray, László Erdodi

This study evaluates the security of web application code generated by Large Language Models, analyzing 2,500 GPT-4 generated PHP websites. These were deployed in Docker containers and tested for vulnerabilities using a hybrid approach of Burp Suite active scanning, static analysis, and manual review. Our investigation focuses on identifying Insecure File Upload, SQL Injection, Stored XSS, and Reflected XSS in GPT-4 generated PHP code. This analysis highlights potential security risks and the implications of deploying such code in real-world scenarios. Overall, our analysis found 2,440 vulnerable parameters. According to Burp's Scan, 11.56% of the sites can be straight out compromised. Adding static scan results, 26% had at least one vulnerability that can be exploited through web interaction. Certain coding scenarios, like file upload functionality, are insecure 78% of the time, underscoring significant risks to software safety and security. To support further research, we have made the source codes and a detailed vulnerability record for each sample publicly available. This study emphasizes the crucial need for thorough testing and evaluation if generative AI technologies are used in software development.

Norbert Tihanyi, Tamas Bisztray, Mohamed Amine Ferrag et al.

This study compares state-of-the-art Large Language Models (LLMs) on their tendency to generate vulnerabilities when writing C programs using a neutral zero-shot prompt. Tihanyi et al. introduced the FormAI dataset at PROMISE'23, featuring 112,000 C programs generated by GPT-3.5-turbo, with over 51.24% identified as vulnerable. We extended that research with a large-scale study involving 9 state-of-the-art models such as OpenAI's GPT-4o-mini, Google's Gemini Pro 1.0, TII's 180 billion-parameter Falcon, Meta's 13 billion-parameter Code Llama, and several other compact models. Additionally, we introduce the FormAI-v2 dataset, which comprises 331 000 compilable C programs generated by these LLMs. Each program in the dataset is labeled based on the vulnerabilities detected in its source code through formal verification, using the Efficient SMT-based Context-Bounded Model Checker (ESBMC). This technique minimizes false positives by providing a counterexample for the specific vulnerability and reduces false negatives by thoroughly completing the verification process. Our study reveals that at least 62.07% of the generated programs are vulnerable. The differences between the models are minor, as they all show similar coding errors with slight variations. Our research highlights that while LLMs offer promising capabilities for code generation, deploying their output in a production environment requires proper risk assessment and validation.

Norbert Tihanyi, Tamas Bisztray, Richard A. Dubniczky et al.

As machine intelligence evolves, the need to test and compare the problem-solving abilities of different AI models grows. However, current benchmarks are often simplistic, allowing models to perform uniformly well and making it difficult to distinguish their capabilities. Additionally, benchmarks typically rely on static question-answer pairs that the models might memorize or guess. To address these limitations, we introduce Dynamic Intelligence Assessment (DIA), a novel methodology for testing AI models using dynamic question templates and improved metrics across multiple disciplines such as mathematics, cryptography, cybersecurity, and computer science. The accompanying dataset, DIA-Bench, contains a diverse collection of challenge templates with mutable parameters presented in various formats, including text, PDFs, compiled binaries, visual puzzles, and CTF-style cybersecurity challenges. Our framework introduces four new metrics to assess a model's reliability and confidence across multiple attempts. These metrics revealed that even simple questions are frequently answered incorrectly when posed in varying forms, highlighting significant gaps in models' reliability. Notably, API models like GPT-4o often overestimated their mathematical capabilities, while ChatGPT-4o demonstrated better performance due to effective tool usage. In self-assessment, OpenAI's o1-mini proved to have the best judgement on what tasks it should attempt to solve. We evaluated 25 state-of-the-art LLMs using DIA-Bench, showing that current models struggle with complex tasks and often display unexpectedly low confidence, even with simpler questions. The DIA framework sets a new standard for assessing not only problem-solving but also a model's adaptive intelligence and ability to assess its limitations. The dataset is publicly available on the project's page: https://github.com/DIA-Bench.

Norbert Tihanyi, Tamas Bisztray, Mohamed Amine Ferrag et al.

Software testing and verification are critical for ensuring the reliability and security of modern software systems. Traditionally, formal verification techniques, such as model checking and theorem proving, have provided rigorous frameworks for detecting bugs and vulnerabilities. However, these methods often face scalability challenges when applied to complex, real-world programs. Recently, the advent of Large Language Models (LLMs) has introduced a new paradigm for software analysis, leveraging their ability to understand insecure coding practices. Although LLMs demonstrate promising capabilities in tasks such as bug prediction and invariant generation, they lack the formal guarantees of classical methods. This paper presents a comprehensive study of state-of-the-art software testing and verification, focusing on three key approaches: classical formal methods, LLM-based analysis, and emerging hybrid techniques, which combine their strengths. We explore each approach's strengths, limitations, and practical applications, highlighting the potential of hybrid systems to address the weaknesses of standalone methods. We analyze whether integrating formal rigor with LLM-driven insights can enhance the effectiveness and scalability of software verification, exploring their viability as a pathway toward more robust and adaptive testing frameworks.

Richard A. Dubniczky, Krisztofer Zoltán Horvát, Tamás Bisztray et al.

Identifying vulnerabilities in source code is crucial, especially in critical software components. Existing methods such as static analysis, dynamic analysis, formal verification, and recently Large Language Models are widely used to detect security flaws. This paper introduces CASTLE (CWE Automated Security Testing and Low-Level Evaluation), a benchmarking framework for evaluating the vulnerability detection capabilities of different methods. We assess 13 static analysis tools, 10 LLMs, and 2 formal verification tools using a hand-crafted dataset of 250 micro-benchmark programs covering 25 common CWEs. We propose the CASTLE Score, a novel evaluation metric to ensure fair comparison. Our results reveal key differences: ESBMC (a formal verification tool) minimizes false positives but struggles with vulnerabilities beyond model checking, such as weak cryptography or SQL injection. Static analyzers suffer from high false positives, increasing manual validation efforts for developers. LLMs perform exceptionally well in the CASTLE dataset when identifying vulnerabilities in small code snippets. However, their accuracy declines, and hallucinations increase as the code size grows. These results suggest that LLMs could play a pivotal role in future security solutions, particularly within code completion frameworks, where they can provide real-time guidance to prevent vulnerabilities. The dataset is accessible at https://github.com/CASTLE-Benchmark.

Rebeka Toth, Tamas Bisztray, Richard Dubniczky

Phishing and spam emails remain a major cybersecurity threat, with attackers increasingly leveraging Large Language Models (LLMs) to craft highly deceptive content. This study presents a comprehensive email dataset containing phishing, spam, and legitimate messages, explicitly distinguishing between human- and LLM-generated content. Each email is annotated with its category, emotional appeal (e.g., urgency, fear, authority), and underlying motivation (e.g., link-following, credential theft, financial fraud). We benchmark multiple LLMs on their ability to identify these emotional and motivational cues and select the most reliable model to annotate the full dataset. To evaluate classification robustness, emails were also rephrased using several LLMs while preserving meaning and intent. A state-of-the-art LLM was then assessed on its performance across both original and rephrased emails using expert-labeled ground truth. The results highlight strong phishing detection capabilities but reveal persistent challenges in distinguishing spam from legitimate emails. Our dataset and evaluation framework contribute to improving AI-assisted email security systems. To support open science, all code, templates, and resources are available on our project site.

Norbert Tihanyi, Bilel Cherif, Richard A. Dubniczky et al.

In this paper, we present the first large-scale study exploring whether JavaScript code generated by Large Language Models (LLMs) can reveal which model produced it, enabling reliable authorship attribution and model fingerprinting. With the rapid rise of AI-generated code, attribution is playing a critical role in detecting vulnerabilities, flagging malicious content, and ensuring accountability. While AI-vs-human detection usually treats AI as a single category we show that individual LLMs leave unique stylistic signatures, even among models belonging to the same family or parameter size. To this end, we introduce LLM-NodeJS, a dataset of 50,000 Node.js back-end programs from 20 large language models. Each has four transformed variants, yielding 250,000 unique JavaScript samples and two additional representations (JSIR and AST) for diverse research applications. Using this dataset, we benchmark traditional machine learning classifiers against fine-tuned Transformer encoders and introduce CodeT5-JSA, a custom architecture derived from the 770M-parameter CodeT5 model with its decoder removed and a modified classification head. It achieves 95.8% accuracy on five-class attribution, 94.6% on ten-class, and 88.5% on twenty-class tasks, surpassing other tested models such as BERT, CodeBERT, and Longformer. We demonstrate that classifiers capture deeper stylistic regularities in program dataflow and structure, rather than relying on surface-level features. As a result, attribution remains effective even after mangling, comment removal, and heavy code transformations. To support open science and reproducibility, we release the LLM-NodeJS dataset, Google Colab training scripts, and all related materials on GitHub: https://github.com/LLM-NodeJS-dataset.