2026-04-02
Rohini Bhosale, Pankaj Chandre, Sushma Mehetre, Swati Powar, Shubhra Mathur, Arun Ghandat
IntroductionThe rapid evolution of artificial intelligence from static large language models to autonomous, agentic AI systems has introduced capabilities such as persistent memory, tool-augmented reasoning, and multi-agent collaboration. While these advancements significantly enhance real-world applicability, they also create a new and underexplored class of privacy risks, including unintended retention, propagation, and amplification of sensitive information across tasks, users, and execution cycles. Existing research predominantly focuses on stateless or single-inference models, leaving the privacy implications of agentic systems insufficiently understood.MethodsThis study presents a comprehensive architectural analysis of data leakage in agentic AI systems. The proposed framework models the end-to-end agent workflow and systematically examines how sensitive information can traverse key components, including persistent memory modules, planning and reasoning processes, tool invocation layers, inter-agent communication channels, and feedback-driven autonomy loops. Based on this architecture, a structured taxonomy of leakage pathways is developed and mapped to realistic threat models and attack vectors observed in practical deployments.ResultsThe analysis identifies multiple leakage pathways unique to agentic AI systems, demonstrating how data can persist, propagate, and be unintentionally exposed across system components and operational cycles. The findings reveal that these leakage mechanisms are more complex and pervasive than those observed in traditional large language model settings, particularly due to the integration of memory, tools, and multi-agent interactions.DiscussionThe study highlights the limitations of existing LLM-centric privacy and security defenses when applied to autonomous agentic systems. It emphasizes the need for lifecycle-aware, component-level mitigation strategies that address privacy risks across the entire agent workflow. The proposed architectural perspective provides a foundation for designing privacy-by-design agentic AI systems and supports safer deployment in sensitive and regulated domains.
DOI: 10.3389/fcomp.2026.18027272026-03-31
Jemimah Nathaniel, Solomon Sunday Oyelere, Jarkko Suhonen, Matti Tedre
Generative artificial intelligence (GenAI) is used in programming education; however, its adoption can introduce pedagogical misalignment, shallow cognitive engagement, and ethical risks that threaten the sustenance of programming skills of students. This study evaluated the GenAI programming education framework’s ability to sustain higher-order thinking skills (HOTS) and programming logic while mitigating pedagogical, cognitive, and ethical barriers in Java programming. A between-group mixed-methods experiment was conducted amongst 124 undergraduate students (62 in the control group and 62 in the experimental group) over 7 weeks. Learning outcomes were assessed using pretests and posttests, analyzed with baseline-adjusted ANCOVA and MANCOVA, and supplemented with trace-based learning analytics from GenAI logs collected at time points (Weeks 3 and 7). The experimental group showed a baseline-adjusted advantage on HOTS (adjusted mean difference = 0.29; p < 0.001; adjusted Hedges’ g = 0.80) and a smaller but significant improvement in programming logic (adjusted mean difference = 0.21; p = 0.047; adjusted Hedges’ g = 0.36), alongside a multivariate group effect across domains. Log-derived indices also showed larger gains in pedagogical alignment and cognitive engagement, reflected in more frequent task decomposition and debugging behaviors. Ethical engagement has also increased, indicating consistent hallucination and data sensitivity awareness. Path modelling indicated that the intervention increased changes in pedagogical, cognitive, and ethical engagement. Pedagogical alignment and cognitive engagement were positively associated with post-test HOTS and programming logic, whereas ethical engagement was negatively associated with HOTS but not significantly associated with programming logic. Overall, the findings suggest that GenAI becomes more educationally beneficial in programming when guided by a structured approach.
DOI: 10.3389/fcomp.2026.17898292026-03-25
Pankaj Chandre, Pallavi Bhujbal, Reetika Kerketta, Jyoti Nandimath, Bhagyashree Shendkar, Rohini Bhosale
IntroductionPhishing attacks have evolved rapidly with the integration of artificial intelligence, posing serious threats to digital trust and cybersecurity. Traditional and AI-assisted phishing techniques still rely on partial human intervention, limiting their adaptability and scalability. Recent advances in agentic artificial intelligence have enabled fully autonomous, goal-driven phishing campaigns capable of planning, personalizing, and executing attacks across multiple communication channels.MethodsThis study investigates the capabilities of agentic AI–enabled phishing by examining its core functional components and operational characteristics. A conceptual architectural perspective is presented to illustrate how autonomous planning, contextual intelligence, multi-modal content generation, and adaptive feedback mechanisms interact to support automated phishing campaigns.ResultsThe analysis demonstrates that agentic AI significantly enhances phishing capabilities by enabling continuous optimization, contextual personalization, and adaptive decision-making during attack execution. The interaction of these architectural components allows phishing systems to dynamically refine strategies and potentially evade conventional detection mechanisms.DiscussionThe study highlights the increasing detection challenges posed by agentic AI–driven phishing systems and examines the associated technical, organizational, and societal risks. Emerging defense strategies and future research directions are discussed to address the evolving threat landscape. Overall, the findings emphasize the urgent need for adaptive and AI-driven countermeasures to effectively mitigate next-generation phishing attacks.
DOI: 10.3389/fcomp.2026.17950452026-03-25
Husheng Pan, Cuiting Kong, Lie Zhang
Addressing current practical challenges in exhibitions in the data-intelligent era—such as an overemphasis on form over content and insufficient emotional resonance—as well as the lack of a systematic theoretical framework to guide practice, this paper draws on embodied cognition and dynamical systems theory to explore the internal mechanisms of interaction design for museum exhibitions in the data-intelligent era from the perspective of cognitive generation. It identifies four core elements of such interaction design, namely the Body Perception Layer, the Body Action Layer, the Environmental Construction Layer and the Meaning Construction Layer, and on this basis constructs a cyclic embodied interaction design framework for museums (the PCAE model) that reveals the dynamic flow of information between visitors and the data-intelligent exhibition environment. Using the data-intelligent interactive exhibit “Dialogue With the Master” at the Confucius Museum in China as a case study, the paper further validates the feasibility and scientific soundness of the proposed framework. This framework introduces a new embodied cross-disciplinary theoretical perspective for research on interaction design in museums in the data-intelligent era and provides an operational design tool that offers designers a clear guiding pathway for optimizing interactive experiences, thereby holding substantial practical value for design practice and theoretical exploration.
DOI: 10.3389/fcomp.2026.1757509