Abstract
The rapid integration of Artificial Intelligence (AI) into cybersecurity has transformed both the offensive and defensive dimensions of cybercrime. AI enables attackers to automate phishing, digital impersonation, malware development, ransomware, credential theft, and other forms of cybercrime at unprecedented speed and scale. At the same time, cybersecurity systems increasingly employ machine learning, deep learning, Natural Language Processing (NLP), behavioral analytics, predictive analytics, and biometric technologies to detect and respond to these threats. However, high reported model accuracy does not necessarily translate into effective real-world protection. Cybersecurity environments are characterized by highly imbalanced threat distributions, false-positive risks, computational constraints, deployment complexity, and substantial differences in access to security technologies between large organizations and individual end users. This paper develops a systems-oriented framework for evaluating AI-enabled cybersecurity beyond accuracy alone. It examines AI cybersecurity technologies according to multiple dimensions, including accuracy, interpretability, computational cost, deployment complexity, false-positive risk, and accessibility. The study further examines two system-level asymmetries: the accessibility gap between enterprise users and individual end users, and the difference between the speed of AI-enabled offensive activity and defensive response. The analysis demonstrates that effective AI cybersecurity should be evaluated as an integrated socio-technical system rather than solely as a high-performing detection model. The paper concludes that cybersecurity protection requires a combination of technically appropriate AI methods, realistic evaluation conditions, human oversight, organizational capacity, and accessible protection mechanisms. This systems perspective provides a basis for evaluating whether AI cybersecurity technologies can deliver meaningful protection to end users in real-world environments.
1. Introduction
Cybercrime refers to malicious activities conducted through computers, digital devices, networks, and the Internet. These activities include phishing, identity theft, hacking, malware, ransomware, credential theft, unauthorized access, and data theft, often occurring in combination. Traditionally, such offenses required substantial human involvement, including the creation and distribution of phishing messages, identification of software vulnerabilities, theft of credentials, and development of malicious software. Although these forms of cybercrime have existed for decades, they continue to cause significant financial, privacy, and security risks to individuals, corporations, and governments. The emergence of Artificial Intelligence (AI), however, has fundamentally altered the technical characteristics of these offenses by making them faster, more sophisticated, scalable, adaptive, and increasingly automated [1]. AI enables attackers to automate activities that previously required extensive human effort while also generating content that is more convincing and difficult for users and conventional security mechanisms to distinguish from legitimate activity [2].
One prominent example is AI-driven phishing, in which generative AI can be used to produce realistic phishing emails, fraudulent messages, deceptive banking websites, and conversational interfaces designed to resemble legitimate services [2]. These capabilities allow attacks to be customized for individual targets while simultaneously being deployed at a much larger scale [3]. Another emerging threat is AI-powered voice cloning and digital impersonation, through which AI systems can reproduce a person’s voice from relatively limited audio samples and combine this capability with publicly available information to facilitate identity-based fraud, financial deception, and social engineering [4]. AI is also increasingly relevant to automated hacking, malware, and ransomware development [5]. AI-assisted systems can identify software vulnerabilities, automate attack processes, and modify malicious behaviour in response to defensive mechanisms, potentially making detection and mitigation more difficult [6]. In the case of ransomware, AI can assist attackers in identifying vulnerable targets and locating sensitive information that may increase the impact of an attack [7]. Similarly, AI-assisted credential theft and identity-based cyber fraud can generate convincing login interfaces, support automated credential attacks, and enhance social-engineering operations [2]. The use of AI for unauthorized access and data theft therefore introduces a broader system-level cybersecurity challenge, because AI systems can analyse large quantities of information and identify potentially valuable or sensitive data at a speed and scale that can exceed conventional human-driven processes [8].
2. Conceptual Findings
The detection of fraud, malware, phishing attacks, and network intrusions by modern-day cybersecurity systems largely depends on the application of artificial intelligence technology. This technology utilizes techniques like machine learning, deep learning, natural language processing, behavioral analysis, and biometric identification to identify such threats in less time than a human analyst would. Deep learning-based detection systems have been achieving accuracy rates between 97.5% and 98.2%; a widely cited Android malware obfuscation detector, AndrODet, has reported accuracy figures of 92.02%, 81.41% and 80.66% across different detection tasks, and hybrid natural language processing models applied to phishing detection have reported accuracy as high a 97%. Several organizations have successfully implemented AI technologies to defend cybersecurity and fraud detection; Darktrace developed an AI-powered autonomous response system for detection and reduction of ransomware attacks at a French hospital group [9]. Similarly, Mastercard uses an AI-based system called DI (Decision Intelligence) to detect fraudulent financial transactions [10].
The Accuracy Illusion—A reported accuracy figure of 97 or 98 per cent invites an intuitive assumption; this assumption does not generally hold, and the reason generally lies in the real-world cybersecurity data. Malicious activity—genuine attacks, intrusions, or fraudulent transactions—occurs far less frequently than original activity within any functioning network. A benchmark dataset used to train and test an AI model contains roughly equal numbers of malicious and benign activities; a real network might contain only one malicious event for every hundred, thousand, or even a million benign ones. Under such conditions, even one small error rate produces a very large number of false alarms; this result is a phenomenon security professionals term “alert fatigue”: the analysts are confronted with an overwhelming volume of false alarms relative to the genuine threats [8].
What makes this worth examining critically is not the fact that the problem exists but how inaccurately it is disclosing cybersecurity reporting. Although some AI technologies, such as Mastercard’s Decision Intelligence, reported significant reductions in false positives [10], many studies only reported high accuracy scores, but these studies do not explicitly specify whether the models were tested using real-world network data or balanced datasets, nor did they report any kind of false-positive rates. Therefore, accuracy should be considered alongside the false-positive rates when evaluating the effectiveness of the cybersecurity AI technologies [8].
Accuracy is an insufficient basis for judging an AI security technique. A related but separate problem is that the existing industry rarely compares different AI techniques against one another using a consistent set of analytics, predictive analytics, biometric authentication—making a direct comparison difficult for a reader trying to decide which approach fits best for the given problem [11].
The observation that can be made is that natural language processing emerges from this index as the most favorable technique overall, as it performs adequately across every criterion rather than just excelling narrowly on one. But what about accessibility? Multinational companies have wide access to the technologies; do common end users such as students, employees, professors, etc. have access to these technologies?
The accessibility Gap between Enterprise and End User—the index above assumes discretion to choose freely among the six techniques; this assumption does not hold for the majority of individuals and small-scale organizations that AI cybersecurity tools are meant to protect. Nearly every well- documented real-world example of these technologies is used by large and well-resourced organizations with a dedicated technical staff along with sufficient budgets that can absorb any complexity and authentication. These documented examples are not incidental; it reflects how the industry can favor and provide for enterprises and leave smaller organizations and the individual end users comparatively unaccounted for.
Small organizations are disproportionately targeted by attackers because they are perceived to carry weaker security infrastructure, and they tend to detect the breaches far slowly compared with large enterprises with stronger security infrastructure. The recurring obstacle appears to be cost, not unawareness of the need for such tools. From an individual perspective, the Indian Citizen Financial Cyber Fraud Reporting and Management System, which has reportedly helped recover approximately ₹8189 crore from 2021 to 2025, is interesting because it demonstrates that an alternative means of access exists to AI-enhanced security than the means exemplified by the enterprise-level approaches described above [12]. In point of fact, the individual end user rarely uses a personal instance of intrusion detection powered by deep learning and behavioral analytics; the protection offered to this individual is derived, if at all, from the efficiency and AI assistance of the reporting system put in use after the crime is committed, as opposed to any AI tools that might assist in preventing the crime beforehand. The contrast drawn here—between the use of AI as a preventative technology accessible mainly to better-funded institutions, and the use of AI as a post hoc remedial measure available to individual users via reporting infrastructure—in this author’s opinion, has been somewhat underestimated where the concept of “AI protecting end users” has been assumed as a unitary one.
The Asymmetry of Offensive and Defensive Times—Another trend becomes apparent when findings relating to speed from AI cybersecurity studies are put side by side, not examined separately. In particular, according to IBM X-Force, an AI-powered phishing email takes about five minutes to generate, whereas it takes about sixteen hours to do so with the aid of a skilled human social engineer—an example of the application of AI for an attack [2]. Additionally, one organization that uses automated AI responses has cut its incident containment time down from three weeks to nineteen minutes—an example of the application of AI for defense [9]. When analyzed separately, both numbers merely demonstrate what the technology can accomplish in its particular field. But put together, these numbers beg the following question: is defensive capacity catching up to the offensive capacity or lagging behind it [11,13]? The following is presented in the Table 1.
Table 1.
Offense–Defense Speed Comparison.
In such a case, the defensive advancement ratio is noticeably larger compared to the offensive ratio, implying a shift in balance towards the defender. Such an assumption should be taken with a grain of salt since the defensive ratio relates to the containment process that takes place only after the detection of the problem, while the offensive ratio is related to the speed of the production of attacks, rather than their speed of detection. It would be inappropriate to conclude that those ratios reflect two symmetrical phases of a similar process. All that can be concluded is that both offensive and defensive actions have been expedited due to the presence of similar technology. At the same time, the question of how much detection itself, as opposed to the response to the detection, has caught up with the speed of generation of the AI-based attacks still needs answering. The comparative analysis of the time required for offensive cyber activities and defensive incident response is presented in Table 2 [14,15,16,17,18].
Table 2.
AI cybersecurity technique suitability Index.
3. Legal Framework
In India, AI governance is applied from a multitude of acts, primarily the Information Technology Act, 2000 (IT Act), the Digital Personal Data Protection Act, 2023 (DPDP Act) and the India AI Governance Guidelines: Enabling Safe and Trusted AI Innovation published by the Ministry of Electronics and Information Technology (MeitY) supplement each other in the prosecution of AI-enabled cybercrimes [19]. The IT Act has criminalized several AI-empowered cybercrimes, and the DPDP Act has placed conditions on the processing of personal data used by AI systems [20].
Further, AI Governance Guidelines introduce a framework to support responsible AI innovation and address the concerns of individuals suffering harms resulting from AI [12]. Together, these legislations are responsible for addressing AI-empowered cybercrimes.
3.1. AI-Generated Phishing
This action is addressed in the Information Technology Act under Section 66D that provides for cheating by personation by using any computer resource [19]. It would also be addressed under Section 66C in cases where an attacker falsely uses the credentials of any other person [19]. The DPDP Act requires that the processing of personal data is conducted lawfully and appropriately safeguarded [20].
The AI Governance Guidelines provide that use of AI-powered tools to engage in cybercrime like phishing is an instance of ‘malicious use of AI’ [12].To combat this threat, the Guidelines propose various technical interventions like content authentication and AI risk assessments [12].
3.2. AI-Powered Voice Cloning and Digital Impersonation
This act squarely falls within the ambit of Section 66C and 66D of the IT Act (identity theft and cheating by personation by use of computer resource, respectively) [12]. Since the recorded voices of individuals or other biometric identifiers constitute ‘personal data’, the processing thereof is subject to the stringent rules and procedures outlined in the DPDP Act [20].It necessitates due diligence on the part of the organizations dealing with such data, while the AI Governance Guidelines recommend greater accountability, increased transparency and better authentication systems [12].
3.3. AI-Generated Automated Malware, Ransomware and Hacking Tools
Under the Information Technology Act, unauthorized access and the introduction of computer programs to damage or disrupt computer systems fall within Section 43(f) (imposing civil liability) and Section 66 (criminalizing the act when carried out dishonestly or fraudulently) [19].
Furthermore, in cases where such malware, ransomware or hacking tools are used to access and exfiltrate data, the DPDP Act will also come into play [20].
The AI Governance Guidelines suggest developing specific risk-assessment frameworks and reinforcing India’s cybersecurity infrastructure, and promoting voluntary standardization in high-risk AI systems [12].
3.4. AI-Aided Credential Theft and Identity-Based Cyber Fraud
In the IT Act, these acts of AI-driven fraud may be criminalized under Section 66C (identity theft) and Section 66D (cheating by personation) [19]. For the organizations handling personal data, the Data Protection Act will hold them accountable for adequately securing the personal data of users and protecting it from unauthorized access and misuse [20]. Under the AI Governance Guidelines, an increase in AI-assisted identity fraud warrants increased awareness, cybersecurity literacy, and strengthened capabilities among the law enforcement agencies to identify and prevent such incidents [11].
3.5. AI-Assisted Data Theft and Unauthorized Disclosure of Confidential Information
Data extraction without authorization may fall within Section 43(f), and the acts may be considered criminal if carried out with dishonest or fraudulent intent under Section 66 [19]. Further, any unauthorized disclosure of confidential information obtained using electronic resources or in breach of lawful contract is covered under Section 72 and 72A of the IT Act [19].
In these cases, the Digital Data Protection Act would impose liability for unlawful processing of personal data, unless the Data Fiduciary establishes adequate security safeguards and protects against data breaches [20].
The AI Governance Guidelines emphasize robust accountability mechanisms, enhanced algorithmic auditing and transparent practices among developers and deployers of AI [12]. The Indian legal framework currently comprises the Information Technology Act, 2000 (IT Act), the Digital Personal Data Protection Act, 2023 (DPDP Act) and the proposed India AI Governance Guidelines which address AI-enabled cybercrimes through criminalization of various cyber offences (IT Act), regulation of the personal data relied upon by AI systems (DPDP Act) and promotion of responsible and safe use of AI technologies (AI Governance Guidelines) [20]. It is noteworthy that the India AI Governance Guidelines conclude that existing Indian legislation like the IT Act and DPDP Act can, in the near term, tackle the concerns arising out of AI-enabled cybercrimes [12]. However, given the rapid evolution and advancement in AI technologies, targeted amendments and specific AI-centric regulatory measures would eventually become inevitable to ensure the safety and security of individuals in this rapidly evolving digital landscape.
4. Discussions and Conclusions
Cybercrime has evolved from relatively direct forms of unauthorized access and fraud into increasingly sophisticated, automated, and scalable attacks. The integration of Artificial Intelligence has accelerated this transformation by enabling attackers to generate convincing phishing content, perform identity-based deception, automate malicious activities, and adapt attacks with reduced human intervention [1]. At the same time, AI has become an important component of cybersecurity systems, supporting threat detection, behavioral analysis, fraud prevention, and automated response [9].
The findings of this study demonstrate, however, that the effectiveness of AI-enabled cybersecurity should not be evaluated through accuracy alone. A reported accuracy of 97% or 98% may appear highly effective in a conventional machine-learning evaluation, but cybersecurity environments are characterized by extreme class imbalance in which malicious events are relatively rare. Under such conditions, even a relatively low false-positive rate can generate a large volume of alerts and contribute to alert fatigue. AI cybersecurity evaluation must therefore incorporate realistic datasets and operational measures such as false-positive rates rather than relying exclusively on headline accuracy figures [8].
The comparative analysis presented in Table 1 further demonstrates that the most accurate AI technique is not necessarily the most suitable cybersecurity solution. Accuracy must be considered alongside interpretability, computational cost, deployment complexity, and false-positive risk. This systems-level approach is particularly important for organizations that lack the technical infrastructure or financial resources required to deploy highly complex AI systems. The suitability of a cybersecurity technology should therefore be determined by the environment in which it will operate rather than by benchmark performance alone [8].
The study also identifies an important accessibility gap. Advanced AI cybersecurity technologies are frequently deployed by large organizations with dedicated technical personnel and substantial resources, while small organizations and individual end users may receive considerably less direct access to comparable capabilities. Consequently, the existence of sophisticated AI cybersecurity technologies should not be equated with universal cybersecurity protection. AI-based protection must also be evaluated according to who can access it, where it is deployed, and whether users possess the resources necessary to operate or benefit from it.
The comparison of offensive and defensive speed provides a further systems-level insight. AI has accelerated both attack generation and defensive response [2]. However, offensive generation and defensive containment represent different stages of the cybersecurity process, and their reported improvement ratios should therefore not be interpreted as directly symmetrical. The more important observation is that AI is compressing the time available for cybersecurity decision-making on both sides. This increases the importance of rapid detection, appropriate automation, human oversight, and resilient response architectures [9].
These findings support a broader conceptualization of AI cybersecurity as an integrated socio-technical system consisting of AI models, data, computational resources, detection mechanisms, human analysts, organizational capabilities, and end users. Within such a system, technical performance is necessary but insufficient. A cybersecurity solution must also be operationally appropriate, economically and computationally feasible, interpretable enough to support decision-making, and sufficiently accessible to the population it is intended to protect [8].
The legal analysis further demonstrates that existing Indian legislation—including the Information Technology Act, 2000, the Digital Personal Data Protection Act, 2023, and the emerging AI governance framework—already provides mechanisms relevant to several forms of AI-enabled cybercrime [4]. These mechanisms address areas including identity theft, cheating by impersonation, unauthorized access, data misuse, and malicious use of AI [19]. Nevertheless, the rapid evolution of AI technologies suggests that cybersecurity governance will need to continue adapting as increasingly autonomous systems become capable of generating, modifying, and executing malicious activity [12].
Accordingly, the principal contribution of this study is a shift from “How accurate is the AI?” to “How effective is the AI-enabled cybersecurity system under realistic conditions?” This distinction is critical. A technically sophisticated model that performs well on a benchmark but generates excessive false alarms, requires inaccessible infrastructure, or cannot be effectively deployed by smaller organizations may provide less practical protection than a simpler and more accessible system [8].
Future research should therefore evaluate AI cybersecurity systems using realistic and highly imbalanced datasets, standardized multi-dimensional evaluation criteria, longitudinal deployment data, and user-centered measures of effectiveness. Further research could also examine adaptive detection architectures, human-AI collaboration, explainable threat detection, lightweight cybersecurity models for small organizations, and AI-based protection mechanisms designed specifically for individual end users [8].
Ultimately, AI has the potential to strengthen cybersecurity substantially, but its value should not be measured by technical capability alone. A technically successful AI cybersecurity model is not necessarily an operationally successful cybersecurity system. Meaningful protection requires the alignment of computational performance, realistic evaluation, deployment feasibility, human expertise, organizational capacity, and accessibility. Only through this systems-oriented approach can AI fulfil its potential as a practical safeguard against the rapidly evolving landscape of AI-enabled cybercrime.
Author Contributions
D.A.A.—Project administration, Supervision, Writing—original draft, H.A. Writing—review & editing, Data curation, H.D. Writing—review & editing, Investigation, M.S.T. Writing—review & editing, Investigation, S.M.K. Writing—review & editing, Validation. All authors have read and agreed to the published version of the manuscript.
Funding
This research received no external funding.
Institutional Review Board Statement
Not applicable.
Informed Consent Statement
Not applicable.
Data Availability Statement
Data are available in this manuscript.
Conflicts of Interest
The authors declare no conflict of interest.
References
- Brooks, C. AI polymorphic threats are forcing a rethink of cybersecurity. Forbes, 21 February 2026. Available online: https://www.forbes.com/sites/chuckbrooks/2026/02/21/ai-polymorphic-threats-are-forcing-a-rethink-of-cybersecurity/ (accessed on 9 July 2026).
- IBM Security. AI vs. Human Deceit: Unravelling New-Age Phishing Tactics. IBM Think. Available online: https://www.ibm.com/think/x-force/ai-vs-human-deceit-unravelling-new-age-phishing-tactics (accessed on 15 July 2026).
- IBM Security. Understanding the Future of Offensive AI in Cybersecurity. IBM Think. Available online: https://www.ibm.com/think/x-force/understanding-future-of-offensive-ai-in-cybersecurity (accessed on 15 July 2026).
- DuckDuckGoose. How AI Voice Cloning Works: 2026 Guide. 2026. Available online: https://www.duckduckgoose.ai/blog/how-ai-voice-cloning-works (accessed on 13 July 2026).
- CybelAngel. Polymorphic Malware: How AI Is Making It Harder to Detect. 8 April 2026. Available online: https://cybelangel.com/blog/ai-threats-polymorphic/ (accessed on 9 July 2026).
- Europol. New 2026 “IOCTA” Highlights Sophisticated Tactics and Emerging Challenges in the Digital Landscape. 1 May 2026. Available online: https://www.europol.europa.eu/media-press/newsroom/news/new-2026-iocta-highlights-sophisticated-tactics-and-emerging-challenges-in-digital-landscape (accessed on 13 July 2026).
- Industrial Cyber. Europol IOCTA 2026 Report Flags Shift to Industrialised Cybercrime Powered by AI, Ransomware and Data Theft. 29 April 2026. Available online: https://industrialcyber.co/reports/europol-iocta-2026-report-flags-shift-to-industrialised-cybercrime-powered-by-ai-ransomware-and-data-theft/ (accessed on 15 July 2026).
- National Institute of Standards and Technology. Artificial Intelligence Risk Management Framework (AI RMF 1.0); NIST AI 100-1; U.S. Department of Commerce: Washington, DC, USA, 2023. Available online: https://nvlpubs.nist.gov/nistpubs/ai/nist.ai.100-1.pdf (accessed on 15 July 2026).
- Darktrace. Major French Hospital Group Stops Ransomware Attack with Darktrace AI. 2026. Available online: https://www.darktrace.com/news/major-french-hospital-group-stops-ransomware-attack-with-darktrace-ai (accessed on 10 July 2026).
- Mastercard. AI is helping banks save millions by transforming payment fraud prevention. Mastercard Newsroom, 6 February 2026. Available online: https://www.mastercard.com/global/en/news-and-trends/Insights/2026/ai-is-helping-banks-save-millions-by-transforming-payment-fraud-prevention.html (accessed on 15 July 2026).
- Help Net Security. Humans Are Still Better than AI at Crafting Phishing Emails, But for How Long? 26 October 2023. Available online: https://www.associationdatabase.com/aws/FRPA/pt/sd/news_article/543777/_PARENT/layout_details/false (accessed on 12 July 2026).
- Ministry of Electronics and Information Technology, Government of India. India AI Governance Guidelines: Enabling Safe and Trusted AI Innovation. 5 November 2025. Available online: https://static.pib.gov.in/WriteReadData/specificdocs/documents/2025/nov/doc2025115685601.pdf (accessed on 15 July 2026).
- VentureBeat. IBM X-Force Pits ChatGPT Against Humans: Who’s Better at Phishing? 2023. Available online: https://venturebeat.com/security/ibm-x-force-pits-chatgpt-against-humans-whos-better-at-phishing (accessed on 12 July 2026).
- Berman, D.S.; Buczak, A.L.; Chavis, J.S.; Corbett, C.L. A Survey of Deep Learning Methods for Cyber Security. Information 2019, 10, 122. [Google Scholar] [CrossRef] [Scilit]
- Salloum, S.; Gaber, T.; Vadera, S.; Shaalan, K. A Systematic Literature Review on Phishing Email Detection Using Natural Language Processing Techniques. IEEE Access 2022, 10, 65703–65727. [Google Scholar] [CrossRef] [Scilit]
- Haroun, K.; Zitouni, A.; Zenakhri, A.; Guessoum, M.A.; Boubchir, L. Efficient Deep Learning for Biometrics: Overview, Challenges and Trends in Ear of Frugal AI. arXiv 2026, arXiv:2602.08809. [Google Scholar]
- Splunk. User and Entity Behavior Analytics (UEBA) Overview in Splunk Enterprise Security. Available online: https://help.splunk.com/en/splunk-enterprise-security-8/administer/8.4/user-and-entity-behavior-analytics (accessed on 12 July 2026).
- Vectra AI. User and Entity Behavior Analytics (UEBA): The Complete Guide to Behavioral Threat Detection. Available online: https://fr.vectra.ai/topics/user-and-entity-behavior-analytics-ueba (accessed on 12 July 2026).
- Information Technology Act, 2000 (India). Available online: https://www.meity.gov.in/static/uploads/2024/03/IT-Act-Rules_2000_0.pdf (accessed on 15 July 2026).
- Digital Personal Data Protection Act, 2023 (India). Available online: https://www.meity.gov.in/static/uploads/2024/06/2bf1f0e9f04e6fb4f8fef35e82c42aa5.pdf (accessed on 15 July 2026).
Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to people or property resulting from any ideas, methods, instructions or products referred to in the content. |
© 2026 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution (CC BY) license.