A new systematic review published in Discover Networks examines the convergence of zero-click attacks and Zero-Trust Architecture (ZTA), highlighting the urgent need for adaptive, AI-driven security frameworks in modern digital infrastructure. The study, authored by Enobong Archibong, Philip Asuquo, and Bliss Stephen, synthesizes 79 high-impact sources using the PRISMA methodology to evaluate how interaction-less exploits bypass traditional perimeter-based defenses. The authors argue that the rapid expansion of cloud-native infrastructure and mobile computing has rendered legacy security models obsolete, necessitating a shift toward resilient frameworks that align with United Nations Sustainable Development Goal 9.
Zero-click attacks represent a sophisticated class of cyber threats that compromise devices without any user interaction. Unlike phishing or one-click attacks, which rely on social engineering or inadvertent user actions, zero-click exploits leverage vulnerabilities in data-processing applications such as instant messengers, email clients, and telephony stacks. These attacks are particularly devastating in mobile ecosystems, where devices are always-on and automatically process rich media in the background, allowing attackers to execute code remotely through specially crafted data packets.
The review identifies the Pegasus spyware and the FORCEDENTRY exploit as prominent examples of this threat class. FORCEDENTRY, for instance, allowed attackers to infect iPhones by sending a hidden PDF file disguised as a GIF, triggering malware execution through the iMessage rendering engine. The study notes that such exploits often target high-privileged system services and media processing libraries, such as the Skia graphics library, to bypass modern operating system mitigations like Address Space Layout Randomization.
In response to these invisible threats, the authors advocate for Zero-Trust Architecture, a security model based on the premise that a breach is inevitable or has already occurred. Formalized in the NIST SP 800-207 standard, ZTA moves security controls from the network edge to individual resources, users, and assets. By mandating continuous authentication, least-privilege access, and micro-segmentation, ZTA seeks to limit the blast radius of an attack, ensuring that even if a device is compromised, the attacker cannot move laterally within the network.
The review details the core architectural components of ZTA, including the Policy Decision Point (PDP) and the Policy Enforcement Point (PEP). The PDP acts as the central authority for access decisions, utilizing a trust algorithm to evaluate risk based on inputs such as user identity, device health, and behavioral attributes. The PEP, situated on the data plane, enforces these decisions by monitoring and controlling connections, effectively acting as a gatekeeper that can terminate sessions if a device’s health status changes or if anomalous behavior is detected.
A significant portion of the review focuses on the integration of Artificial Intelligence and Machine Learning into ZTA to address the operational burden of real-time trust decisions. The authors highlight the SmartTrust framework as a primary case study, which utilizes a hybrid deep learning approach combining Convolutional Neural Networks (CNNs), Long Short-Term Memory (LSTM) networks, and Reinforcement Learning. This architecture enables real-time, context-aware risk scoring, allowing the system to detect subtle anomalies in network traffic that traditional rule-based systems might miss.
Empirical evaluations of the SmartTrust framework on the UNSW-NB15 benchmark dataset demonstrated a detection accuracy of up to 99.27% for data breaches and 99.19% for insider threats. The framework achieved an average detection latency of approximately 120 milliseconds, a critical metric for real-time enforcement in high-speed cloud infrastructures. Additionally, the study reports that the hybrid model reduces false alarms by over 40% compared to standalone CNN or Random Forest models, enhancing its practical utility in enterprise environments.
Despite these promising results, the review acknowledges several limitations and challenges in deploying AI-driven ZTA. The high accuracy rates are largely derived from training on standardized benchmarks, which may not fully capture the noise and polymorphic variance of live network traffic. The authors warn of overfitting risks, where models may struggle to generalize to novel, zero-day exploits that utilize previously unseen techniques. Furthermore, the increasing prevalence of encrypted traffic, such as TLS 1.3, limits the ability of the PDP to perform deep packet inspection without introducing significant decryption overhead.
The study also addresses the vulnerability of AI models to adversarial machine learning attacks. Sophisticated adversaries can utilize adversarial perturbations to trick the system into assigning low-risk scores to malicious traffic, or employ model poisoning to inject clean labels for malicious behaviors during the retraining phase. To mitigate these risks, the authors propose the development of Robust ZTA, which integrates adversarial training and defensive distillation techniques to harden the policy engine against such evasion strategies.
Explainability remains a critical barrier to the adoption of AI-driven ZTA in regulated sectors such as healthcare and finance. Black-box neural networks provide high accuracy but low transparency, making it difficult for human administrators to explain why access was denied. The review proposes a Dual-Path Explainable AI Architecture, where a fast path handles sub-150ms allow/deny decisions, while an asynchronous path generates explanations using frameworks like SHAP and LIME. This approach aims to balance the need for real-time enforcement with the interpretability required for compliance audits.
The authors also explore emerging threats posed by Generative AI, such as the EchoLeak vulnerability, which allows data exfiltration through invisible prompt injection in emails. To defend against these AI-native threats, the review suggests the implementation of Semantic Firewalls that operate at the intent layer of communication. By utilizing Natural Language Processing to analyze the context of data payloads, these firewalls can detect logic bombs or prompt injections hidden within legitimate application traffic, providing a layer of protection that traditional stateful inspection cannot offer.
Finally, the review outlines a strategic roadmap for future research, emphasizing the need for privacy-preserving machine learning techniques like Federated Learning and Differential Privacy. These methods would allow policy engines to train on decentralized device telemetry without exposing raw user data, creating a zero-knowledge security layer. By addressing the scalability of behavioral analysis and the resilience of AI models against adversarial evasion, the study aims to foster more resilient information and communication technology infrastructure essential for modern economic stability.
The review’s systematic methodology involved a multi-phased literature search across major digital repositories, including IEEE Xplore, ACM, Springer, and ScienceDirect, covering the period from 2010 through 2026. The search strategy utilized Boolean operators to isolate literature at the intersection of threat, architectural, and intelligent defence domains. From an initial identification of 245 records, the authors removed 40 duplicates and screened 205 records based on title and abstract. After excluding 86 articles that did not meet technical criteria for zero-interaction exploits or ZTA architectural depth, 119 reports were sought for retrieval. The final synthesis comprises 79 studies, ensuring the technical analysis is grounded in validated, peer-reviewed research rather than anecdotal evidence. This rigorous filtering process was designed to minimise selection bias and ensure technical depth in the evaluation of both exploit mechanics and defensive architectures.
Zero-click attacks are technically complex and expensive to develop, which typically restricts their use to nation-states or elite hacking groups targeting high-value individuals such as journalists, activists, or executives. The review maps these attacks to specific tactics within the MITRE ATT&CK Mobile framework, noting that they bypass traditional security awareness training and often leave no visible forensic footprint on the device interface. A key distinction in the threat landscape is the shift from user-dependent vectors to interaction-less exploits that target autonomous background processes. For example, the 2019 WhatsApp vulnerability allowed spyware installation simply by calling a target, where the mere ringing data packet was sufficient to execute code. Similarly, a 2023 Outlook exploit enabled the theft of password hashes by sending a calendar invitation that the application processed automatically to set a reminder sound, demonstrating the breadth of vulnerable system components.
Modern mobile operating systems have implemented several hardware-assisted mitigations to raise the cost of zero-click development. These include the Memory Tagging Extension (MTE), found in ARMv8.3 and ARMv9 architectures, which detects and prevents memory safety violations by tagging memory allocations and checking them at runtime. Pointer Authentication Codes (PAC) cryptographically sign pointers to prevent Return-Oriented Programming techniques used in sophisticated exploits. Additionally, syscall isolation mechanisms like Seccomp and GWP-ASan restrict the system calls a media-parsing process can make, thereby limiting the blast radius of a compromised parser. On iOS, the BlastDoor service provides a dedicated, sandboxed, and highly restricted environment written in Swift, a memory-safe language, specifically designed to parse incoming iMessage payloads. These measures represent a move beyond simple sandboxing toward hardware-rooted security to counter the sophistication of modern Advanced Persistent Threats.
The review highlights a documented sophistication gap between academic detection frameworks and real-world APT toolkits. Most academic research is trained on static datasets that do not account for the Just-In-Time obfuscation used by groups like NSO Group. Academic models often assume the exploit exists in a network packet, whereas real-world zero-clicks may reside in the uninitialized memory of a Graphics Processing Unit or an Image Decoder. In contrast, real-world APT toolkits such as KingsPawn bypass behavioural analytics by using living-off-the-land binaries, which are legitimate system tools used to perform malicious tasks. This makes them indistinguishable from standard system noise to most AI models. The review notes that while static analysis provides a first line of defence against known vulnerabilities, it remains highly susceptible to code packing and polymorphism, necessitating a shift toward robust behavioural monitoring integrated with ZTA.
Network enforcement patterns play a critical role in neutralizing zero-click threats that bypass the user interface. The review identifies Software-Defined Networking (SDN) driven micro-segmentation as a primary pattern, where the Policy Enforcement Point can provision micro-perimeters around individual application workloads. If a media parsing library is compromised, the SDN controller can instantly revoke flow rules for that specific container, preventing lateral movement. In 5G-enabled environments, ZTA leverages Network Slicing to isolate traffic from compromised User Equipment, ensuring that a compromise in a public-facing slice cannot bridge into the secure enterprise environment. Furthermore, Identity-Based Routing tags every packet with a session-specific, cryptographically signed metadata token issued by the Policy Administrator. Routers and switches acting as PEPs only forward traffic with valid tokens, rendering the internal network dark to unauthorised or malformed packet attempts typical of zero-click reconnaissance.
Subject of Research: Review of zero-click attacks and zero-trust security model: concepts, architecture, state of the art, and future directions
Article Title: Review of zero-click attacks and zero-trust security model: concepts, architecture, state of the art, and future directions
Article References: Archibong, E., Asuquo, P., & Stephen, B. (2026). Review of zero-click attacks and zero-trust security model: concepts, architecture, state of the art, and future directions. Discover Networks, 2(1), Article 10. https://doi.org/10.1007/s44354-026-00023-4
Image Credits: AI Generated
DOI: 10.1007/s44354-026-00023-4
Keywords: Review, zero-click, attacks, zero-trust, security, model, concepts, architecture, state, future, directions, scientific research
News Source: Denise Maddox. (November 12, 2026). Review Links Zero-Click Exploits to AI-Driven Zero-Trust Security. Scienmag.
