AI is rewriting social engineering, from phishing defenses to autonomous machine deception

Artificial intelligence has quietly moved beyond being a tool that learns, predicts, and decides. According to a new editorial published in Frontiers in Artificial Intelligence, AI has become an infrastructure that mediates how human behavior is observed, interpreted, and influenced — and that shift is fundamentally redefining social engineering, the strategic exploitation of human behavior and cognitive vulnerabilities to sway decisions or induce actions. The editorial, written by Vasile Daniel Păvăloaia and colleagues from institutions in Romania, Spain, Italy, Poland, and Albania, synthesizes ten contributions to a Research Topic examining AI’s role in social engineering across business, healthcare, and industrial settings in Europe, Asia, Africa, and Latin America.

The central argument is that the question facing organizations is no longer whether a person can be deceived by a single malicious message. Instead, the authors contend, the real challenge is how AI systems shape what is visible, how people and situations are represented, and how decisions get made. Across the ten studies, a consistent pattern emerges: AI capabilities expand organizational efficiency and analytical precision while simultaneously generating novel vectors of behavioral vulnerability and strategic influence. The editorial organizes this dynamic as a conceptual progression, from observing and interpreting information, through predicting behavior, to persuasion, decision support, and ultimately autonomous action.

One of the clearest vantage points on this intersection comes from a bibliometric review of AI in phishing detection by Popescul and Radu, which analyzed 1,096 studies. The review documents a clear thematic shift from traditional machine learning classifiers toward deep learning, natural language processing, hybrid models, and classifier stacking. Yet despite these technical advances, the study’s most striking conclusion is that human psychological vulnerability remains the primary attack surface. Technical detection tools and user-oriented awareness training, the authors emphasize, are complementary safeguards rather than substitutes for one another — a finding that undercuts the hope that better algorithms alone can close the phishing gap.

Other contributions map how AI transforms the informational foundations of influence itself. Through topic modeling applied to large datasets, Karlsson and Dalipi found that discourse around surveillance is dominated by practical and military applications, which receive far more attention than abstract privacy or ethical concerns. Their empirical findings also highlight the bi-directional nature of contemporary surveillance, reflecting a shift in how influence and information are distributed. In a related vein, Soomro and Baeshen demonstrate through customer big data analytics that behavioral information can become organizational value when analytics are connected to customer satisfaction and performance. Chen and colleagues, applying machine learning to consumer choices, identified satisfaction, perceived benefits, and understanding of relevant “ingredients” as keys to behavioral segmentation — showing that some human motivations are more predictable to computational systems than others.

These studies trace a progression from observing behavior to making it increasingly legible to machines. But the editorial argues that the future of social engineering and influence will be defined less by traditional predictive models and reactive detection than by what the authors call Agentic AI: architectures, multimodal systems, synthetic identities, deepfakes, and multi-agent orchestration frameworks capable of executing continuous, adaptive interaction loops without human intervention. Unlike conventional prompt-driven uses of large language models, agentic systems possess goal-directed autonomy. They can plan multi-step workflows, execute API calls, negotiate, and transact on behalf of individuals and organizations. As a consequence, the editorial predicts, social engineering may increasingly extend beyond human-to-human interaction toward machine-to-machine deception and machine-to-human influence mediated by autonomous proxies.

Generative AI illustrates why these developments resist a simple benefit-versus-risk framing, because the same features that support legitimate decision-making can enable deception, persuasion, and social engineering influence. In recruitment, Tripathi and colleagues document that the integration of large language models is expanding rapidly, with a heavy focus on efficiency metrics — even though empirical validation of control mechanisms remains limited despite high awareness of ethical risks such as algorithmic bias. A multi-stakeholder study by Jonnala and colleagues goes further, identifying interconnected risk relationships rather than isolated threats. Toxic outputs emerge as a major driver from the developer perspective, with causal links to improper usage and hallucinations. The study also describes feedback loops in which harmful outputs can reinforce bias and trust deficits — dynamics particularly relevant to social engineering, because the potential for influence depends not on a single malicious act but on the interaction of technical outputs, human responses, and organizational safeguards.

The combined evidence suggests that emerging vulnerabilities do not arise from AI capabilities alone, but from how those capabilities interact with human responses, organizational safeguards, and decision contexts. Looking ahead, the editorial points to algorithmic advances in large language models — specifically real-time hallucination suppression and self-correcting reasoning — as potentially critical for making generative AI more reliable in AI-assisted decision-making. At the same time, other contributions highlight genuine benefits. De Micco and colleagues examine embodied AI in disability and rehabilitation, where it can promote independence, autonomy, and social participation, though the authors warn against a purely functionalist approach that could neglect dignity, human relationships, and equitable access. Shahidi Hamedani and colleagues connect AI in business operations with operational efficiency, sustainability, personalization, and decision support, while emphasizing data quality, privacy, security, AI literacy, employment change, and the digital divide.

AI adoption itself can become part of the social engineering transformation, and its effects are highly contextual. Evidence from companies in an emerging economy, reported by Perez-Cepeda and colleagues, extends the Technology Acceptance Model by identifying a reciprocal relationship between perceived usefulness and perceived ease of use — suggesting an interaction in which experience, learning, and perceived value reinforce one another rather than adoption being a one-way response to technological usability. In a different setting, Kyubwa examines informal economies characterized by limited infrastructure, where AI acts as a conditional capability amplifier, contributing to a phenomenon of compressed professionalization among youth who use these tools to bypass formal institutional pathways. The editorial stresses that this amplification remains uneven because infrastructure, skills, sustained access, and institutional conditions all matter — a perspective that guards against a technologically deterministic reading of AI adoption and underscores the importance of human decision-making in shaping technological outcomes. Future research, the authors suggest, should investigate how decentralized agentic AI, locally adaptive autonomous systems, and multimodal AI acting on behalf of users can be designed to safeguard human autonomy and equitably amplify localized capabilities across diverse socio-economic infrastructures.

Synthesizing the ten papers, the authors conclude that the field of social engineering stands on the verge of a structural shift. Traditionally, social engineering has been conceptualized around human actors designing deceptive interventions to influence other people. Emerging AI architectures are changing that dynamic, and the editorial proposes three interconnected research questions to guide the field: when AI-driven personalization crosses the boundary from assistance into social engineering; how AI-generated identities affect trust; and how meaningful human oversight can be ensured as autonomous systems proliferate. The editorial also suggests that managing the boundary between legitimate AI-mediated influence and manipulation will require future oversight mechanisms, algorithmic audits, and behavioral watermarking to verify persuasive intent.

The editorial’s closing framing is deliberately balanced: AI is neither an inherently protective shield nor an unmitigated threat, but a transformative infrastructure reshaping human autonomy, trust, and influence. As AI systems become more autonomous, agentic, and embedded in consequential decisions, the authors argue, responsibility for their effects must remain traceable across the actors involved in their design, deployment, and use. Meaningful human control, they caution, requires more than formal human oversight — decision-makers must be able to understand, critically assess, and, where necessary, challenge or override AI-supported outputs. As an editorial synthesizing a Research Topic, the work reflects the authors’ interpretation of the ten studies rather than new empirical findings of its own, but its diagnosis is clear: the next paradigm of social engineering will be written not by humans alone, but by autonomous machines acting on their behalf.

Subject of Research: How artificial intelligence reshapes social engineering, its impacts and ethics across business, healthcare, and industry

Article Title: Editorial: Integrating AI in social engineering: impacts and ethics across business, medicine, and industry

Article References: Păvăloaia, V. D., Georgescu, M. R., Pozo, F., Maris, B. M., Sulikowski, P., Domnori, E., & Greavu-Serban, V. (2026). Editorial: Integrating AI in social engineering: impacts and ethics across business, medicine, and industry. Frontiers in Artificial Intelligence, 9, Article 1987515. https://doi.org/10.3389/frai.2026.1987515

Image Credits: AI Generated

DOI: 10.3389/frai.2026.1987515

Keywords: artificial intelligence, social engineering, Agentic AI, Generative AI, large language models, phishing detection, surveillance, algorithmic bias, human autonomy, AI ethics, deepfakes, AI adoption