Security and Risk Implications of Transformer-Based Large Language Models

Phishing attacks continue to grow in scale and sophistication, causing substantial financial losses and privacy breaches worldwide. Recent advances in large language models (LLMs) have brought significant changes to the generation and detection of phishing content. This study systematically investigates the dual role of LLMs in facilitating phishing attacks and strengthening countermeasures. Using the PRISMA methodology, authors screened 142 records published between January 2023 and April 2025 and identified 36 eligible studies from major academic databases, including IEEE Xplore, ScienceDirect, ACM Digital Library, Web of Science, and Scopus. A comprehensive and rigorous analysis was conducted of research trends/themes over time, dataset characteristics, and the LLM architectures/models employed. The findings reveal that most studies relied on manually generated datasets rather than publicly available benchmark datasets, and that GPT-based models received considerably more attention than other LLM architectures. The review demonstrates that LLMs substantially enhance the generation of phishing content by producing coherent, contextually relevant, and persuasive email and website content. This capability lowers the technical barrier for attackers and potentially increases attack effectiveness. Conversely, LLMs also strengthen defensive strategies by enabling more effective analysis of textual and visual content for phishing detection. In many cases, LLM-based approaches outperform traditional machine learning and deep learning methods and, in certain contexts, approach or match human-level performance. Overall, the findings suggest that LLMs have accelerated and automated phishing-related processes, simultaneously intensifying the threat landscape and advancing defensive capabilities. • The first in-depth study to review LLMs usage in Phishing attack generation and detection. • The study reveals LLMs have accelerated and automated phishing-related processes, elevating both threats and defence mechanisms. • GenAI-based multimodal phishing attacks are on the rise due to the wider adoption of GenAI tools in general.

Transformer-based large language models (LLMs), such as ChatGPT-4, are increasingly used to streamline clinical practice, of which radiology reporting is a prominent aspect. However, their performance in interpreting complex anatomical regions from MRI data remains largely unexplored. This study investigates the capability of ChatGPT-4 to produce clinically reliable reports based on orbital MR images, applying a multimetric, quantitative evaluation framework in 25 patients with orbital lesions. Due to inherent limitations of current version of GPT-4, the model was not fed with MR volumetric data, but key 2D images only. For each case, ChatGPT-4 generated a free-text report, which was then compared to the corresponding ground-truth report authored by a board-certified radiologist. Evaluation included established NLP metrics (BLEU-4, ROUGE-L, BERTScore), clinical content recognition scores (RadGraph F1, CheXbert), and expert human judgment. Among the automated metrics, BERTScore demonstrated the highest language similarity, while RadGraph F1 best captured clinical entity recognition. Clinician assessment revealed moderate agreement with the LLM outputs, with performance decreasing in complex or infiltrative cases. The study highlights both the promise and current limitations of LLMs in radiology, particularly regarding their inability to process volumetric data and maintain spatial consistency. These findings suggest that while LLMs may assist in structured reporting, effective integration into diagnostic imaging workflows will require coupling with advanced vision models capable of full 3D interpretation.Supplementary InformationThe online version contains supplementary material available at 10.1038/s41598-025-19669-1.

Recently, large language models (LLMs) have attracted considerable attention due to their remarkable capabilities. However, LLMs’ generation of biased or hallucinatory content raised significant concerns, posing major challenges for their practical application. Many studies have dedicated efforts to address these critical issues, adopting various approaches to mitigate bias and hallucinations in LLM-generated content. Remarkably, no review papers have synthesized insights on these two primary problems. Addressing this gap, this paper aims to conduct a simultaneous and dual-focused review of the current landscape of research. The discussions encompass widely used and newly proposed benchmarks and evaluation methods on bias and hallucination in LLMs. This paper also investigates advanced mitigation methods and present a taxonomy based on different mitigation strategies. Moreover, a comparative analysis of the sources, mitigation methods, and evaluation methods for bias and hallucination is included. In the end, this paper provides a synthesis of current research trends and suggests potential directions for future research to address bias and hallucination in LLMs, considering the ongoing challenges in this field.

This article sets in with the question whether current or foreseeable transformer-based large language models (LLMs), such as the ones powering OpenAI’s ChatGPT, could be language users in a way comparable to humans. It answers the question negatively, presenting the following argument. Apart from niche uses, to use language means to act. But LLMs are unable to act because they lack intentions. This, in turn, is because they are the wrong kind of being: agents with intentions need to be autonomous organisms while LLMs are heteronomous mechanisms. To conclude, the article argues, based on structural aspects of transformer-based LLMs, that these LLMs have taken a first step away from mechanistic artificiality to autonomous self-constitution, which means that these models are (slowly) moving into a direction that someday might result in non-human, but equally non-artificial agents, thus subverting the time-honored Kantian distinction between organism and mechanism.

In recent years, Transformer-based large language models (LLMs) have significantly improved upon their text generation capability. Mental health is a serious concern that can be addressed using LLM-based automated mental health counselors. These systems can provide empathetic responses to individuals in need while considering the negative beliefs, stigma, and taboos associated with mental health issues. Considering the large size of these LLMs makes it difficult to deploy these automated counselors on low cost/resource devices such as edge devices. Therefore, the motivation of the present study to analyze the effectiveness of lightweight LLMs in the development of automated mental health counseling systems. In this study, lightweight open source LLMs such as Google’s T5s (small variant), BARTB (base variant), FLAN-T5s (small variant), and Microsoft’s GODELB (base variant) have been fine-tuned for automated mental health counseling task utilizing a diverse set of datasets publicly available online. The experimental results reveal that BART’s base variant outperformed the other models across all key metrics such as ROUGE-1, ROUGE-2, ROUGE-L, and BLEU with scores of 0.4727, 0.2665, 0.3554, and 25.3993 respectively. In comparison to other models, BART-base model generated empathetic, and emotionally supportive responses. These findings highlight the potential of lightweight LLMs (small size LLMs), in advancing the field of LLM-based mental health counseling solutions and underscore the need for exploration of lightweight LLMs for this mental health counseling use case. The code for this work is available at the following link: https://github.com/diviitmg03/Comparative-analysis-of-LLMs-.git.

UNSTRUCTURED Misleading information has significant implications for society but can have disastrous impact for health matters. Transformative artificial intelligence (AI) tools such as large language models (LLMs) have the potential for limitless content generation (including fake), soon making internet information impossible to assess using traditional human approaches. We asked if the same LLMs (GPT4 and Gemini1-5-Pro) could be part of a more scalable solution. We tested 2 publicly available LLMs for their ability to identify misinformation in HealthReleases previously labeled by human experts. We found that simple prompts lead to overall low accuracy (F1 Macro 0,45 (GPT4) and 0,49 (Gemini1-5Pro)), but very different profiles for each LLM. Adding specific criteria used by experts to critically assess the Releases enhanced Gemini (0.66) but surprisingly reduced GPT4 (0,37) performances. We therefore developed a novel approach incorporating summaries of expert feedback into prompts and then observed major improvements in performance for both LLMs(GPT4;0.63 and Gemini1-5Pro; 0.96). Our study provides the first use case of LLMs as high throughput proofing of medical text, but more importantly provides insights into LLMs’ “truth biases”. We provide a novel paradigm integrating knowledge into the prompts which may reduce the need for LLM training, and the requirement for ever larger datasets and compute power. Importantly, we show how experts could and need to be involved in LLMs used to enhance their performance and potentially minimize the data wall issue.

Video generation using Large Language Models (LLMs) has shown promising potential, effectively leveraging the extensive LLM infrastructure to provide a unified framework for multimodal understanding and content generation. However, these methods face critical challenges, i.e., token redundancy and inefficiencies arising from long sequences, which constrain their performance and efficiency compared to diffusion-based approaches. In this study, we investigate the impact of token redundancy in LLM-based video generation by information-theoretic analysis and propose Vision Representation Compression (VRC), a novel framework designed to achieve more in both performance and efficiency with less video token representations. VRC introduces learnable representation compressor and decompressor to compress video token representations, enabling autoregressive next-sequence prediction in a compact latent space. Our approach reduces redundancy, shortens token sequences, and improves model's ability to capture underlying video structures. Our experiments demonstrate that VRC reduces token sequence lengths by a factor of 4, achieving more than 9~14x acceleration in inference while maintaining performance comparable to state-of-the-art video generation models. VRC not only accelerates the inference but also significantly reduces memory requirements during both model training and inference.

The recent developments in artificial intelligence (AI), particularly in light of the impressive capabilities of transformer-based Large Language Models (LLMs), have reignited the discussion in cognitive science regarding whether computational devices could possess semantic understanding or whether they are merely mimicking human intelligence. Recent research has highlighted limitations in LLMs’ reasoning, suggesting that the gap between mere symbol manipulation (syntax) and deeper understanding (semantics) remains wide open. While LLMs overcome certain aspects of the symbol grounding problem through human feedback, they still lack true semantic understanding, struggling with common-sense reasoning and abstract thinking. This paper argues that while adding sensory inputs and embodying AI through sensorimotor integration with the environment might enhance its ability to connect symbols to real-world meaning, this alone would not close the gap between syntax and semantics. True meaning-making also requires a connection to subjective experience, which current AI lacks. The path to AGI must address the fundamental relationship between symbol manipulation, data processing, pattern matching, and probabilistic best guesses with true knowledge that requires conscious experience. A transition from AI to AGI can occur only if it possesses conscious experience, which is closely tied to semantic understanding. Recognition of this connection could furnish new philosophical insights into longstanding practical and philosophical questions for theories in biology and cognitive science and provide more meaningful tests of intelligence than the Turing test.

Large language models (LLMs) are becoming a significant source of content generation in social networks, which is a typical complex adaptive system (CAS). However, due to their hallucinatory nature, LLMs produce false information that can spread through social networks, which will impact the stability of the whole society. The uncertainty of LLMs false information spread within social networks is attributable to the diversity of individual behaviors, intricate interconnectivity, and dynamic network structures. Quantifying the uncertainty of false information spread by LLMs in social networks is beneficial for preemptively devising strategies to defend against threats. To address these challenges, we propose an LLMs hallucination-aware dynamic modeling method via agent-based probability distributions, spread popularity and community affiliation, to quantify the uncertain spreading of LLMs hallucination in social networks. We set up the node attributes and behaviors in the model based on real-world data. For evaluation, we consider the spreaders, informed people, discerning and unwilling non-spreaders as indicators, and quantified the spreading under different LLMs task situations, such as QA, dialogue, and summarization, as well as LLMs versions. Furthermore, we conduct experiments using real-world LLM hallucination data combined with social network features to ensure the validity of the proposed quantifying scheme.

Introduction Game theory has long served as a foundational tool in cybersecurity to test, predict, and design strategic interactions between attackers and defenders. The recent advent of Large Language Models (LLMs) offers new tools and challenges for the security of computer systems. In this work, we investigate whether classical game-theoretic frameworks can effectively capture the behaviors of LLM-driven actors and bots. Methods Using a reproducible framework for game-theoretic LLM agents, we investigate two canonical scenarios—the one-shot zero-sum game and the dynamic Prisoner's Dilemma—and we test whether LLMs converge to expected outcomes or exhibit deviations due to embedded biases. We experiments on four state-of-the-art LLMs and five natural languages (English, French, Arabic, Vietnamese, and Mandarin Chinese) to assess linguistic sensitivity. Results For both games, we observe that the final payoffs are influenced by agents characteristics such as personality traits or knowledge of repeated rounds. We also uncover an unexpected sensitivity of the final payoffs to the choice of languages, which should warn against indiscriminate application of LLMs in cybersecurity applications and call for in-depth studies, as LLMs may behave differently when deployed in different countries. We also employ quantitative metrics to evaluate the internal consistency and cross-language stability of LLM agents. Discussion In addition to uncovering unexpected behaviors requiring attention by scholars and practitioners, our work can help guide the selection of the most stable LLMs and optimizing models for secure applications.

The integration of large language models (LLMs) into medical diagnostics has garnered substantial attention due to their potential to enhance diagnostic accuracy, streamline clinical workflows, and address health care disparities. However, the rapid evolution of LLM research necessitates a comprehensive synthesis of their applications, challenges, and future directions. This scoping review aimed to provide an overview of the current state of research regarding the use of LLMs in medical diagnostics. The study sought to answer four primary subquestions, as follows: (1) Which LLMs are commonly used? (2) How are LLMs assessed in diagnosis? (3) What is the current performance of LLMs in diagnosing diseases? (4) Which medical domains are investigating the application of LLMs? This scoping review was conducted according to the Joanna Briggs Institute Manual for Evidence Synthesis and adheres to the PRISMA-ScR (Preferred Reporting Items for Systematic Reviews and Meta-Analyses extension for Scoping Reviews). Relevant literature was searched from the Web of Science, PubMed, Embase, IEEE Xplore, and ACM Digital Library databases from 2022 to 2025. Articles were screened and selected based on predefined inclusion and exclusion criteria. Bibliometric analysis was performed using VOSviewer to identify major research clusters and trends. Data extraction included details on LLM types, application domains, and performance metrics. The field is rapidly expanding, with a surge in publications after 2023. GPT-4 and its variants dominated research (70/95, 74% of studies), followed by GPT-3.5 (34/95, 36%). Key applications included disease classification (text or image-based), medical question answering, and diagnostic content generation. LLMs demonstrated high accuracy in specialties like radiology, psychiatry, and neurology but exhibited biases in race, gender, and cost predictions. Ethical concerns, including privacy risks and model hallucination, alongside regulatory fragmentation, were critical barriers to clinical adoption. LLMs hold transformative potential for medical diagnostics but require rigorous validation, bias mitigation, and multimodal integration to address real-world complexities. Future research should prioritize explainable artificial intelligence frameworks, specialty-specific optimization, and international regulatory harmonization to ensure equitable and safe clinical deployment.

This paper discusses a multimodal AI system applied to legal reasoning for tax law. The results given here are very general and apply to systems developed for other areas besides tax law. A central goal of this work is to gain a better understanding of the relationships between LLMs (Large Language Models) and automated theorem-proving methodologies. To do this, we suppose (1) two cases for the theorem-proving system: one where it has a countable number of total meanings for its countable number of atoms and the other is where it has an uncountable number of total meanings for its countable number of atoms, and (2) LLMs can have an uncountable number of token meanings. With this in mind, the results given in this paper use the downward and upward Löwenheim–Skolem theorems and logical model theory to contrast these two AI modalities. One modality focuses on syntactic proofs and the other focuses on logical semantics based on LLMs. Particularly, one modality uses a rule-based first-order logic theorem-proving system to perform legal reasoning. The objective of this theorem-proving system is to provide proofs as evidence of valid legal reasoning when enacted laws are applied to particular situations. These proofs are syntactic structures that can be presented in the form of narrative explanations of how the answer to the legal question was determined. The second modality uses LLMs to analyze and transform a user’s tax query so this query can be sent to a first-order logic theorem-proving system to perform its legal reasoning function. The main goal of our application of LLMs is to enhance and simplify user input and output for the theorem-proving system. Using logical model theory, we show how there can exist an equivalence between laws represented in logic of the theorem-proving system, fixed in time when the theorem-proving system was set up, and new semantics given by LLMs. These results are based on logical model theory and Löwenheim–Skolem theorems.

The application of deep learning to tabular data remains an ongoing challenge, with tree-based models such as XGBoost consistently outperforming neural network methods in most real-world scenarios. Recent advances in Large Language Models (LLMs) highlight their ability to generalize to new tasks via few-shot or one-shot prompting, yet questions remain about their utility in structured, tabular domains—particularly for resource-efficient Small Language Models (SLMs). In this work, we systematically evaluate the effectiveness of both large and small transformer-based language models for tabular classification tasks using few-shot strategies. Our approach investigates input serialization schemes and prompt engineering to maximize performance in low-data regimes. Experiments on benchmark datasets—including Diabetes, Heart Failure, and German Credit Risk—demonstrate that well-designed SLMs can approach, and occasionally match, the performance of much larger models, substantially reducing the need for extensive labeled data and retraining. However, further advances are required for language models to consistently rival the strongest tree-ensemble baselines. Our findings support the idea that SLMs, when properly prompted, offer a promising, flexible, and label-efficient alternative for automating and dynamizing machine learning pipelines on tabular data.

The vigorous development of artificial intelligence (AI) represented by large language models (LLMs) has rapidly promoted the updating and development of educational technology. Agentic workflows (AWs) built based on LLMs can realize complex tasks in the field of education, which allows the emergence of swarm intelligence (SI) through multi-agent collaboration. This study introduces the Agent4EDU (agent for education) framework, which outlines 4 application models in education from the two dimensions of degree of agency and degree of interaction, including human-AI collaboration, AI assistant, instruction execution, and general type. The proposed Agent4EDU framework discusses the paradigm of educational applications of AI agents and promotes the development of the field of AI for education.

The work is dedicated to the study of fundamental prompting techniques to improve the efficiency of using large language models (LLMs). Significant attention is given to the issue of prompt engineering. Various techniques are examined in detail: zero-shot prompting, feedback prompting, few-shot prompting, chain-of-thought, tree of thoughts, and instruction tuning. Special emphasis is placed on Reaction & Act Prompting and Retrieval Augmented Generation (RAG) as critical factors in ensuring effective interaction with LLMs. The features of applying these techniques and their impact on results are highlighted. However, leveraging their full potential requires a careful approach and consideration of application specifics. A review of the parameters of large language models, such as temperature, top P, maximum number of tokens, stop sequences, frequency and presence penalties, etc., is provided. It is noted that prompt development is an iterative process that involves sequential testing of different options to achieve optimal results. All techniques discussed in the study are supported by illustrative examples with obtained results. It is indicated which types of tasks each technique is more suitable for. The study results include comparisons of both fundamental techniques and more advanced technologies such as ReAct and RAG. Prompt engineering is a key technology for the effective use of large language models. It is relevant due to the increasing application of artificial intelligence in all areas of human activity, and its role will only grow with the development of technology. The ability to correctly formulate prompts is becoming an important skill necessary for working with modern large models, especially given their versatility and complexity. Keywords: large language models, prompt engineering, prompting technique, content generation.