Secure Multi-LLM Agentic AI and Agentification for Edge General Intelligence by Zero-Trust: A Survey

TL;DR

This survey explores the integration of zero-trust security principles into multi-LLM agentic AI systems for edge general intelligence, addressing critical security vulnerabilities and proposing a comprehensive framework.

Contribution

It is the first systematic survey of zero-trust security mechanisms applied to multi-LLM systems in edge general intelligence environments.

Findings

Identifies key security risks in multi-LLM systems.

Categorizes zero-trust security approaches into model- and system-level.

Provides a roadmap for future research in zero-trust multi-LLM security.

Abstract

Agentification serves as a critical enabler of Edge General Intelligence (EGI), transforming massive edge devices into cognitive agents through integrating Large Language Models (LLMs) and perception, reasoning, and acting modules. These agents collaborate across heterogeneous edge infrastructures, forming multi-LLM agentic AI systems that leverage collective intelligence and specialized capabilities to tackle complex, multi-step tasks. However, the collaborative nature of multi-LLM systems introduces critical security vulnerabilities, including insecure inter-LLM communications, expanded attack surfaces, and cross-domain data leakage that traditional perimeter-based security cannot adequately address. To this end, this survey introduces zero-trust security of multi-LLM in EGI, a paradigmatic shift following the ``never trust, always verify'' principle. We begin by systematically analyzing the security risks in multi-LLM systems within EGI contexts. Subsequently, we present the vision of a zero-trust multi-LLM framework in EGI. We then survey key technical progress to facilitate zero-trust multi-LLM systems in EGI. Particularly, we categorize zero-trust security mechanisms into model- and system-level approaches. The former and latter include strong identification, context-aware access control, etc., and proactive maintenance, blockchain-based management, etc., respectively. Finally, we identify critical research directions. This survey serves as the first systematic treatment of zero-trust applied to multi-LLM systems, providing both theoretical foundations and practical strategies.

Figures

Tables

Table 1. Table 1. Summary of related surveys

Ref.	Overview	Single-LLM	Multi-LLM	EGI	Zero-Trust
(Das et al., 2025)	A survey comprehensively analyzing security and privacy vulnerabilities of LLMs, including prompt hacking, adversarial attacks, and privacy leakage, alongside mitigation strategies.	✓	✗	✗	✗
(Aguilera-Martínez and Berzal, 2025)	A structured survey categorizing LLM attacks by lifecycle stages, detailing prevention-based and detection-based defenses.	✓	✗	✗	✗
(Friha et al., 2024)	A survey examining security vulnerabilities specific to edge intelligence scenarios involving LLM deployments, highlighting proactive and reactive defenses under resource constraints.	✓	✗	✓	✗
(Gan et al., 2024)	A survey proposing a taxonomy categorizing threats of LLMs by origin (inputs, model flaws, interactions) and consequences (security, privacy, ethical issues), discussing relevant defenses.	✓	✗	✗	✗
(Liu et al., 2024)	A survey providing guidelines for evaluating LLM alignment with security and ethical standards, covering dimensions such as reliability, fairness, and robustness.	✓	✗	✗	✗
(Ko et al., 2025)	A survey categorizing security challenges in multi-LLM systems with a focus on cross-domain interactions, highlighting risks like dynamic grouping and collusion.	✗	✓	✗	✗
(Kong et al., 2025)	A survey analyzing communication security within multi-LLM systems, detailing vulnerabilities and mitigation methods across different interaction patterns and stages.	✗	✓	✗	✗
(Luo et al., 2025)	A survey exploring trust, orchestration, and resource scheduling challenges in multi-LLM EGI deployments.	✗	✓	✓	✗
(Peigne et al., 2024)	A systematic evaluation of security-efficiency trade-offs in collaborative multi-LLM systems, reviewing infectious malicious prompts and targeted defense methods.	✗	✓	✗	✗

Table 2. Table 2. Security issues in multi-LLM systems for EGI

Category	Type of Security Issue	Description
Intra-LLM	Jailbreaks & Prompt Injection	Adversaries craft malicious prompts to bypass safety guardrails, manipulating the LLM’s internal state to generate harmful outputs (Xu et al., 2024; Liao and Sun, 2024).
	Unpredictable Emerging Abilities	Unexpected emergent behaviors in LLMs pose risks due to unforeseen capabilities, such as autonomously generating exploit codes (Fang et al., 2024).
	Data Leakage & Privacy	LLMs inadvertently disclose sensitive information through memorized training data or user-provided confidential inputs (Carlini et al., 2021; Mireshghallah et al., 2022).
	Toxic or Misaligned Responses	LLMs produce biased, offensive, or incorrect content due to training data biases or misalignment with intended operational scenarios, leading to real-world risks (Sahoo et al., 2024).
Inter-LLM	Expanded Attack Surface	Each additional LLM increases entry points for attackers, creating a cascading risk where a compromised model affects multiple peers (Lee and Tiwari, 2024; Shen et al., 2025).
	Over-Permissive Integration	LLMs tightly integrated with privileged system components may unintentionally trigger sensitive operations, resulting in unauthorized actions or escalations of privilege (Zhang et al., 2025; OpenAI et al., 2023).
	Insecure Communication	Communication channels lacking encryption or authentication become vulnerable to prompt injections, impersonation, and eavesdropping (Lee and Tiwari, 2024; He et al., 2025; Zhang et al., 2024).
	Consensus Manipulation	Malicious or compromised LLMs manipulate consensus protocols by injecting false information or dishonest voting, disrupting collaborative decisions (Zhang et al., 2025).
	Cross-Context Data Leakage	Collaboration among LLMs leads to unintended data exposure when sharing context-specific information, potentially violating global privacy policies (Juneja et al., 2025).

Table 3. Table 3. Perimeter-based security vs zero-trust security for multi-LLM systems

Dimension	Sub-dimension	Traditional Perimeter Security	Zero-Trust Security
Core Philosophy	Basic Concept	Expanding trust perimeter in multi-LLM	Building a multi-LLM environment without trust
	Trust Assumption	Components within perimeter are implicitly trusted	No component and entity should be unconditionally trusted at any time
	​​ Technical Enablers (Surveyed in Sections 3.3, 5, and 6)	• Enhance LLM capabilities against attacks (DP, adversarial training, RLHF) • Operate on trusted hardware (TEE) • Set security perimeters (firewalls, RBAC) • Establish trusted communication links	• Strong identification and authentication • Context-aware access control • Stateless and Ephemeral LLM • Proactive maintenance • Distributed management • Micro-segmentation and Isolation • Intelligent monitoring and failure management
Implementation	Deployment	Can independently use individual defense mechanisms, e.g., adversarial training or DP.	Systems engineering requiring multi-mechanism coordination
	Implementation Risk	Failure of one protection mechanism may compromise overall defense	If only implementing partial zero-trust principles while ignoring others, may create serious vulnerabilities in a zero-trust environment. Example: Only doing identity verification without continuous monitoring allows attackers to persist after initial access
Resource Overhead	Training Overhead	High: For instance, RLHF fine-tuning requires extensive human annotation; Adversarial training needs massive attack samples; DP training increases computational complexity	Relatively Low: Mainly relies on existing model capabilities; No large-scale retraining required
	Operational Overhead	Medium: Periodic security rule updates; Hardware maintenance costs	High: Continuous authentication increases computational burden; Real-time monitoring requires substantial computing resources; Dynamic permission management increases system complexity; Multi-system coordination operational complexity
Applicable Scenarios	Environment Features	Relatively closed environments; Simple threat models; Single LLM deployment; Resource-constrained scenarios	Distributed multi-LLM systems; EGI environments; High-security requirement scenarios; Complex threat environments; Cross-domain LLM collaboration
	Typical Applications	Enterprise internal AI assistants, closed-domain QA systems	Autonomous driving systems, smart healthcare networks, multi-institutional collaborative AI
	Target Attacks in EGI	• External penetration attacks (network intrusions, malicious external access, known threat patterns identified by the firewall) • Static threats (predefined attack patterns, signature-based malware, fixed attack vectors)	• Lateral movement (authorized user malicious behavior, privilege abuse, inter-LLM malicious propagation) • Dynamic & complex attacks (prompt injection, jailbreaking, adversarial attacks, model poisoning) • Emerging & unknown threats (emergent capabilities, multi-modal attacks, context-dependent threats)
Advantages		Relatively simple implementation; Clear single-point optimization effects; High technology maturity; Controllable costs	Comprehensive security assurance; Adaptable to dynamic threats; Fine-grained control; Systematic protection
Limitations		Once perimeter is breached, internal systems fully exposed; Difficult to handle insider threats; Static protection mechanisms; High lateral movement risks	Extremely high implementation complexity; Expensive operational costs; High technical team requirements; May impact system performance

Table 4. Table 4. Model-based approaches of zero-trust multi-LLM

	Introduction	Methods	Representative Works
Strong LLM Identity, Authentication, and Authorization	Implements “never trust, always verify” principle by establishing continuous authentication and cryptographic identity verification for each LLM in EGI environments.	Multi-factor Authentication	Adaptive MFA (Zhou et al., 2025): Dynamically adjusts verification requirements based on real-time risk assessments and agent behavior patterns.
			Context-aware MFA (Wang and Wang, 2023): Balances security and usability by tightening protocols during detected anomalies using contextual information.
		Reputation-based Authentication	LLMChain (Nagarajan et al., 2024): Maintains reputation scores based on historical outputs, policy adherence, and user feedback for access control.
			Blockchain Reputation (Bouchiha et al., 2024): Provides transparent, verifiable, and immutable reputation histories using blockchain technology.
		Token-based Authentication	Fine-grained Token Control (Chan, 2025): Issues ephemeral, cryptographically secure tokens encoding specific permissions for LLM interactions.
			Ephemeral Token Management (Joshi, 2025): Enforces principle of least privilege through continuous re-authentication and token expiration.
Context-aware Access Control	Implements “least privilege” principle by context-aware permission management that grants the minimal necessary access rights to LLMs.		AgentSafe (Mao et al., 2025): Protects multi-LLM systems through hierarchical data management with ThreatSieve and HierarCache components.
			EPEAgents (Shi et al., 2025): Minimizes data exchange by providing each LLM only task-relevant information through context-aware filtering.
			ABE-based Access Control (Xiao et al., 2025): Uses Attribute-Based Encryption with policy hiding to ensure only authorized LLMs can decrypt sensitive information.
			Collaborative Memory Framework (Rezazadeh et al., 2025): Encodes memory access permissions as time-evolving bipartite graphs with context-aware policies.
Stateless and Ephemeral LLM Management	Embodies “assume breach” principle through eliminating persistent state and creating disposable LLMs in EGI and isolated execution contexts.	Stateless Management	PagedAttention (Kwon et al., 2023): Implements process-like isolation with dynamic memory allocation and copy-on-write semantics for request isolation.
			vAttention (Prabhu et al., 2025): Advances stateless security through hardware-level isolation using CUDA virtual memory APIs for protected address spaces.
			BlockLLM (Arshad and Halim, 2025): Extends stateless principles through component-level micro-segmentation with cryptographically isolated blocks.
		Ephemeral Management	Self-destructing Models (Henderson et al., 2023): Embed algorithmic time locks that degrade model performance when adapted for harmful tasks.
			Serverless Deployment (Fu et al., 2023): Realizes automatic lifecycle management with security checkpoints and anomaly-triggered termination.

Table 5. Table 5. The representative proactive maintenance approaches to realize zero-trust multi-LLM in EGI

Methods	Description	Representative Works
Intelligent Input Checking	Preemptively analyzes and verifies user prompts to detect and neutralize malicious instructions or jailbreak attempts before LLMs generate harmful content.	LLM Prompt Detection System (Kim et al., 2024): Employs regular expressions and fine-tuned LLMs to identify PII, malicious codes, URLs, and prompt injection commands.
		DefensiveTokens (Chen et al., 2025a): Embeds special tokens into user prompts to disrupt malicious patterns and prevent injection attacks.
		JailGuard (Zhang et al., 2025): Creates multiple mutated variants of user prompts to defend against prompt-based attacks through pattern disruption.
		Layered Filtering Framework (Muliarevych, 2024): Uses GPT-4 as attack validator to achieve 97% detection accuracy for prompt injection attacks.
		SecurityLingua (Li et al., 2025): Employs security-aware prompt compression to highlight critical security instructions and remove distracting context.
Reputation Schemes	Evaluates LLM through reputation to proactively isolate suspicious models before causing damage.	LLMChain Framework (Bouchiha et al., 2024): Evaluates LLM reliability based on previous interactions to identify potentially malicious LLMs proactively.
		PsySafe (Zhang et al., 2024): Employs psychology-based framework to assess LLM behaviors and psychological states for early malicious tendency identification.
Topology-aware Proactive Maintenance	Examines network-wide LLM interaction patterns and topological configurations to predict and prevent misinformation propagation throughout the system.	MedSentry (Chen et al., 2025b): Implements comprehensive evaluation pipeline that benchmarks risk in representative multi-LLM topologies for early vulnerability identification.
		NetSafe (Yu et al., 2025): Examines topological aspects of LLM interactions to determine how network configurations influence misinformation propagation.
		G-Safeguard (Wang et al., 2025): Uses temporal graph modeling to detect anomalies in LLM interactions and disrupt harmful content spread.
		Guardian (Zhou et al., 2025): Leverages topology-guided graph neural networks to proactively identify and isolate malicious LLMs in collaborative environments.

Full text

Multi-LLM, zero-trust, edge general intelligence, agentic AI, security

††copyright: acmlicensed††journalyear: 2025††doi: XXXXXXX.XXXXXXX††journal: CSUR††journalvolume: 9††journalnumber: 9††article: 35††publicationmonth: 9††copyright: none††ccs: General and reference Surveys and overviews††ccs: Networks Mobile networks††ccs: Computing methodologies Artificial intelligence††ccs: Security and privacy Network security

1. Introduction

1.1. Background

Multi-Large Language Model (multi-LLM) systems (Yang et al., 2024) are rapidly emerging as a transformative paradigm in agentic AI (Acharya et al., 2025) and agentification (Gunduz et al., 2025), where multiple specialized LLMs collaborate to solve complex tasks beyond the capabilities of individual agents. This architectural shift is particularly crucial for advancing Edge General Intelligence (EGI) (Chen et al., 2024; He et al., 2025), since different LLMs trained on diverse domain-specific datasets and serving users within certain application domains can be strategically deployed to provide low-latency, customized, and ubiquitous intelligence services. Multi-LLM systems have gained widespread deployment across EGI domains. For instance, Waymo deploys multiple specialized LLMs directly on vehicle chips, where perception LLMs manage sensors, planning LLMs make real-time navigation decisions, and control LLMs execute driving maneuvers111https://waymo.com/blog/2024/10/introducing-emma. Moreover, AWS’s Bedrock implements dynamic LLM routing. For example, in online shopping systems, lightweight LLMs can handle simple product searches and order tracking locally, while complex queries, such as product comparisons, can be routed to powerful cloud LLMs, reducing response latency by up to 50% and operational costs by 30%222https://aws.amazon.com/blogs/machine-learning/multi-LLM-routing-strategies-for-generative-ai-applications-on-aws/.

1.2. Motivations

However, multi-LLM systems expose significant security vulnerabilities that threaten the integrity of EGI deployments. First, each individual LLM remains susceptible to jailbreaking attacks, where adversaries craft malicious prompts to bypass safety guardrails and extract harmful outputs (Xu et al., 2024). Furthermore, the collaborative nature of multi-LLM architectures introduces unique attack surfaces: inter-LLM communication channels can be exploited through prompt injection, where compromised outputs from one LLM cascade as malicious inputs to others (Liu et al., 2023). Additionally, the presence of potentially malicious LLMs poses risks, as one rogue LLM can poison the entire decision-making pipeline (Chen et al., 2024). Moreover, multi-LLM systems blur the boundaries between data ownership and model accountability. When multiple LLMs process sensitive information collaboratively, it becomes challenging to ensure consistent privacy policies across heterogeneous LLMs and trace data leakage sources (Ko et al., 2025). The user-system boundary is similarly obscured, as LLMs in multi-LLM intelligent systems serve dual roles, both as user interfaces and as autonomous agents executing system operations, which undermines traditional access control mechanisms (Ko et al., 2025).

To this end, researchers have focused on improving the trustworthiness and security of multi-LLM systems. For instance, during the pre-training phases, adversarial training (Yu et al., 2024) and Differential Privacy (DP) techniques (Charles et al., 2024) can be employed to strengthen LLMs’ resistance against malicious inputs and prevent models from memorizing users’ private data. During edge deployment, LLMs can be deployed within Trusted Execution Environments (TEEs) on devices, thereby isolating attackers and ensuring that LLM operations remain tamper-proof (Yu et al., 2022). In addition, reactive output filtering mechanisms and firewall-based defenses have been developed to intercept and sanitize potentially harmful content before it reaches end users (Zeng et al., 2024).

Nonetheless, these measures are fundamentally grounded in traditional and intuitive perimeter-based security paradigms, where they construct and expand security boundaries within multi-LLM EGI scenarios to establish trusted zones. Although these approaches demonstrated effectiveness, they face increasingly severe challenges in dynamic multi-LLM deployments. First, LLM capabilities are advancing at an unprecedented pace, often exceeding designers’ expectations, making it difficult to define clear security boundaries. For example, GPT-4 demonstrated unexpected autonomous problem-solving capabilities during safety testing when it successfully bypassed a CAPTCHA challenge by recruiting a human worker on TaskRabbit (OpenAI et al., 2023). Furthermore, many perimeter-based security defenses exhibit reactive characteristics. For instance, with reactive output filtering, the transient existence of malicious content can still cause information leakage and security risks, which is detrimental to EGI (He et al., 2025). Finally, the collaborative nature of multi-LLM systems creates extensive lateral movement opportunities across different security domains, potentially compromising data integrity and undermining the effectiveness of boundary-based protections (NIS, [n. d.]).

Recently, zero-trust security has emerged as a transformative paradigm, gaining widespread attention and achieving significant success across diverse domains, including computer networks, 6G communications, and IoT systems (Cao et al., 2025). Unlike perimeter-based approaches that operate under implicit trust assumptions and respond only after security breaches occur, zero-trust security adopts the foundational principle of “never trust, always verify,” establishing a security-first framework where trust is explicitly earned rather than assumed. Specifically, the integration of zero-trust security in multi-LLM and the corresponding benefits can be explained as follows.

• •

Explicit Verification: No component in the multi-LLM system possesses inherent trustworthiness. Every data access request, tool invocation, and inter-LLM communication should undergo rigorous validation before execution (NIS, [n. d.]). Each user input and LLM output is treated as potentially malicious and should be verified through formal verification. This comprehensive verification scheme specifically addresses the single-point failure vulnerabilities of multi-LLM systems, where a single compromised component can undermine the entire EGI. Moreover, unlike traditional perimeter-based approaches that implicitly trust components within established boundaries, explicit verification ensures continuous authentication even for previously verified entities, effectively preventing the lateral movement attacks that plague collaborative LLM architectures.

• •

Least Privilege: Traditional approaches assume that enhancing LLM capabilities inherently improves service experience and security. However, zero-trust enforces that user requests should be fulfilled using LLMs with the minimum necessary capabilities (Cao et al., 2025), ensuring optimal resource allocation and security posture. Access to data repositories, external APIs, and computational resources is granted dynamically based on specific task requirements and validated trust levels, enabling precise capability matching while maintaining security boundaries (Cao et al., 2025). This intelligent access control prevents over-privileged operations that could expose sensitive resources, while simultaneously optimizing computational efficiency in resource-constrained edge environments.

• •

Continuous Monitoring: Finally, zero-trust LLM maintains persistent surveillance of system activities, assuming that security threats can emerge at any time from any component (Poirrier et al., 2025). Each deviation from established operational patterns triggers immediate security assessments and potential containment measures. This proactive monitoring capability enables early threat detection and rapid response in multi-LLM EGI deployments, maintaining the operational continuity of EGI services. Unlike reactive security measures that respond only after attacks have been taken, continuous monitoring provides real-time visibility into inter-LLM communications and collaborative decision-making processes, enabling the detection of subtle anomalies such as consensus manipulation and cross-context data leakage before they can propagate throughout the distributed system.

1.3. Related Surveys

1.3.1. Single-LLM Security

Recent research into the security and privacy issues of LLMs has generated a broad range of surveys, each contributing unique insights into this complex domain. Initially, Das et al. (Das et al., 2025) presented a comprehensive survey focusing explicitly on the security and privacy challenges of LLMs, analyzing vulnerabilities, such as prompt hacking and adversarial attacks, and privacy attacks such as gradient leakage and membership inference. They highlighted these risks in various applications, providing detailed mitigation strategies and emphasizing future research directions. Building upon this foundation, Aguilera-Martínez et al. (Aguilera-Martínez and Berzal, 2025) extended the discussion by categorizing LLM attacks according to their occurrence at different lifecycle stages, including pre-training, inference, etc. Additionally, they provided a structured overview of defense strategies, separating them into prevention-based and detection-based mechanisms. Furthermore, Friha et al. (Friha et al., 2024) explored the integration of LLMs within edge intelligence, examining the unique architectural challenges, resource constraints, and security vulnerabilities presented when deploying powerful LLMs in resource-constrained environments. Particularly, they categorized defense strategies into proactive measures, which aim to prevent potential threats, and reactive measures, designed to mitigate risks after threats have emerged. Gan et al. (Gan et al., 2024) conducted a comprehensive survey of security risks in LLM-based agents, proposing an innovative taxonomy based on threat sources and impacts. Specifically, they categorized threats by their origins (problematic inputs, model flaws, or input-model interactions) and their consequences (security/safety, privacy, or ethical issues), thus comprehensively discussing security vulnerabilities and defense mechanisms across different components of LLM-based agent architectures. Lastly, Liu et al. (Liu et al., 2024) offered guidelines for evaluating the alignment of LLMs with security and ethical standards. Their survey covered critical dimensions such as reliability, safety, fairness, resistance to misuse, interpretability, adherence to social norms, and robustness. Despite their thoroughness, these surveys collectively reveal common limitations. Primarily, there is insufficient coverage of multi-LLM systems, where interactions between multiple LLMs can introduce complex, emergent security and ethical risks not captured by analyses focusing solely on single-LLM deployments.

1.3.2. Multi-LLM Security

To this end, few works have begun explicitly discussing the multi-LLM scenario. Ko et al. (Ko et al., 2025) categorized security challenges specifically in cross-domain multi-LLM systems, highlighting novel threats arising from inter-domain interactions such as unvetted dynamic grouping and collusion control. Kong et al. (Kong et al., 2025) presented a comprehensive survey on the communication security of multi-LLM systems, outlining security risks across user-LLM, LLM-LLM, and LLM-environment interactions, along with detailed countermeasures for each communication stage. Luo et al. (Luo et al., 2025) explored the architecture, trust, and orchestration challenges of multi-LLM deployments in EGI, emphasizing the importance of dynamic orchestration and resource scheduling to maintain security and efficiency. Furthermore, Peign et al. (Peigne et al., 2024) systematically evaluated the trade-offs between security and collaborative efficiency within multi-LLM systems, introducing concepts such as infectious malicious prompts and reviewing targeted defense strategies. However, these surveys primarily discuss traditional security approaches aimed at enhancing the trustworthiness of LLMs themselves and their generated content, including methods such as adversarial training, LLM pruning, and cryptographic protections. They did not consider the emerging zero-trust paradigm, in which implicit trusts are eliminated in every entity by default. We observe that many advanced studies (Chan, 2025; Shi et al., 2025; Xiao et al., 2025; Arshad and Halim, 2025; Chen et al., 2025a, b; Chong et al., 2025; Wang et al., 2024) have proposed defenses aligning with zero-trust principles, while a comprehensive survey and tutorial systematically discussing multi-LLM security from a zero-trust perspective have yet to emerge. Our survey fills this critical gap (see Table 1).

1.4. Our Contributions and Survey Structure

To the best of our knowledge, we are the first to survey the concept of “zero-trust multi-LLM” in EGI scenarios. We begin by comprehensively introducing essential background knowledge, including the basics of LLM technologies, the fundamentals of multi-LLM architectures, zero-trust security principles, and the emerging paradigm of EGI. Subsequently, we systematically analyze the security challenges faced by multi-LLM systems in EGI deployments. We then survey traditional security approaches based on perimeter security models. Particularly, we analyze the inherent limitations of these approaches and demonstrate how zero-trust principles offer transformative opportunities. Observing that although substantial literature has presented security defenses that align with zero-trust concepts, a comprehensive zero-trust framework for multi-LLM systems remains absent. Hence, we further present a unified vision of zero-trust multi-LLM architectures for EGI. Then, we conduct an extensive survey of zero-trust security mechanisms categorized into model- and system-level approaches. Finally, we identify some critical future research directions that will drive the advancement of zero-trust multi-LLM systems. The main contributions of this survey can be summarized as follows.

• •

We comprehensively introduce the current development status and advantages of multi-LLM systems over traditional single LLMs. Afterward, we comprehensively analyze the critical role that multi-LLM plays in constructing future EGI architectures. Moreover, we introduce the concept and applications of zero-trust security.

• •

We analyze security risks faced by multi-LLM systems in EGI deployments, categorizing threats at both intra-LLM and inter-LLM levels. Subsequently, we summarize traditional perimeter-based security approaches, analyze their limitations, and explore the integration of zero-trust security with multi-LLM architectures, providing a comparative analysis between traditional trustworthy approaches and zero-trust paradigms.

• •

We present a comprehensive zero-trust multi-LLM framework oriented to EGI environments, demonstrating how zero-trust principles can be systematically implemented across distributed multi-LLM deployments. Moreover, we conduct an extensive survey of state-of-the-art zero-trust security mechanisms for multi-LLM systems, categorizing approaches into model- and system-level approaches.

• •

We identify and discuss three critical future research directions that require immediate attention from the research community. These research directions collectively advance the theoretical foundations and practical deployment of zero-trust multi-LLM systems, encouraging further research in this emerging field.

The remainder of this survey is organized as shown in Fig. 1. First, Section 2 provides fundamental background knowledge, covering the basics of LLMs, multi-LLM systems, zero-trust security principles, and their applications in EGI scenarios. Then, Section 3 systematically analyzes the security challenges faced by multi-LLM systems in EGI deployments, categorizing vulnerabilities at both intra-LLM and inter-LLM levels, and surveys traditional trustworthy approaches along with their inherent limitations. Section 4 introduces a comprehensive zero-trust multi-LLM framework for EGI, demonstrating the systematic implementation of zero-trust principles through detailed architectural design and operational workflows. Sections 5 and 6 present an extensive review of zero-trust security mechanisms. The identification and discussion of critical future research directions are presented in Section 7. Finally, Section 8 concludes this survey.

2. Fundamentals and Preliminaries

2.1. Multi-LLM Systems

LLMs represent a revolutionary advancement in AI, evolving from early transformer architectures to sophisticated systems capable of natural language understanding, reasoning, and generation across diverse domains (Cha, [n. d.]). Recent developments have witnessed remarkable progress from GPT-3’s 175 billion parameters to GPT-4’s multimodal capabilities, demonstrating unprecedented performance in Q&A, code generation, image drawing, etc. (Kirillov et al., 2023).

Nonetheless, a single LLM often faces notable limitations in complex real-world scenarios. First, each LLM is fundamentally constrained by performance ceilings and lacks specialization, making it less effective in handling complex, multi-step tasks that require decomposition, domain expertise, or long-term planning (Wu et al., 2024). Additionally, single LLMs create computational bottlenecks and represent single points of failure, lacking fault tolerance, load distribution, and horizontal scaling capabilities (Hong et al., 2024). Most critically, single LLMs cannot engage in collaborative reasoning, cross-validation, and consensus mechanisms that significantly improve accuracy and reduce hallucinations (Du et al., 2024). These structural limitations highlight the need for collaborative intelligence in practical deployments.

In contrast, a multi-LLM system refers to an architecture in which multiple LLMs are organized to cooperate on solving one complex task (Li et al., 2024). This approach addresses the fundamental limitations of single LLM systems by enabling specialized model deployment for domain-specific expertise, distributed computational load across multiple instances, and collaborative reasoning mechanisms that enhance performance (Wu et al., 2024; Li et al., 2024). Moreover, multi-LLM architectures provide inherent fault tolerance through redundancy and horizontal scalability through intelligent routing. For instance, in autonomous driving systems, a multi-LLM system can deploy specialized LLMs for perception (i.e., processing sensor data), planning (i.e., route optimization and decision-making), and control (i.e., vehicle dynamics), where each LLM contributes its expertise while collaborating through standardized communication protocols to achieve reliable autonomous navigation. As shown in Fig. 2(left), the technical enablers of multi-LLM systems are as follows.

• •

Orchestration: As shown in Fig. 2(left), current multi-LLM systems exhibit diverse topologies, including linear pipelines, star-shaped patterns with a central coordinator, hierarchical trees, etc. (Zhang et al., 2025). Linear structures fit staged workflows with sequential steps, while star and tree topologies enable central control or layered decomposition. The orchestration mechanisms manage task delegation, scheduling, and output integration. Systems may use explicit controller agents, such as Claude’s lead agent333https://www.anthropic.com/engineering/built-multi-agent-research-system, or implicit workflows via prompt chaining. Moreover, orchestration defines agent profiles and interfaces, aligning the multi-LLM system with external user intent.

• •

Routing and LLM Selection: Efficient routing mechanisms determine which LLMs and tools are called for a given task. Static routing relies on predefined rules, while adaptive routing leverages prompts or learned policies. For instance, the MRKL framework (Karpas et al., 2022) allows an LLM to dynamically select from expert modules for specific operations. Industrial platforms like AWS Bedrock444https://aws.amazon.com/cn/bedrock/ also support LLM routing and load balancing based on task requirements and system states.

• •

Shared Memory & State: Effective collaboration in multi-LLM systems requires mechanisms for consistent memory access and contextual alignment across agents. Memory can be categorized into several types, including short-term conversational buffers, long-term memory stored in vector databases, and shared domain knowledge (Li et al., 2024). For example, MetaGPT (Hong et al., 2024) uses a shared message pool to coordinate agent behaviors asynchronously. Similarly, LLMs may access structured external knowledge sources via Retrieval-Augmented Generation (RAG) techniques. This shared memory not only provides coherence in context management but also supports task continuity, enables strategy adaptation from prior episodes, and promotes consistency in goal-oriented reasoning.

• •

Communication Protocols: LLMs should communicate using well-defined formats and structured interaction patterns (Ehtesham et al., 2025). Communication may occur through natural language dialogue, structured JSON exchanges, or symbolic representations. Protocol designs span synchronous turn-taking, asynchronous message broadcasting, and debate-based deliberations evaluated by a coordinator. Emerging protocol standards, such as Model Context Protocol (MCP) (Kong et al., 2025), Agent Communications Protocol (ACP), and Agent Network Protocol (ANP) (Ehtesham et al., 2025), support both inter-LLM communication and interactions between agents and external tools, ensuring interoperability and consistent multi-agent behavior.

2.2. Multi-LLM for EGI

2.2.1. The Vision of EGI

EGI is an emerging paradigm that extends edge computing toward more general-purpose AI capabilities (Chen et al., 2024). Traditional edge computing shifts data processing from centralized cloud infrastructures to decentralized edge nodes near data sources, thereby reducing latency and network congestion (Du et al., 2022). Early edge intelligence enabled real-time AI services localized at the edge, such as object detection by smart surveillance cameras (Chen et al., 2024), but such systems are confined to a single, pre-defined task or application domain. EGI takes a further step forward by drawing inspiration from artificial general intelligence (Bubeck et al., 2023), aiming to endow edge devices with human-like cognitive versatility. Specifically, EGI targets edge systems that can perform comprehension, reasoning, decision-making, and adaptation across diverse scenarios, even those not seen during training (He et al., 2025).

A key enabler of this shift is the emergence of LLMs, which serve as general-purpose cognitive engines trained on vast corpora of multimodal data. Furthermore, multi-LLM systems are particularly significant for realizing EGI. By orchestrating multiple specialized LLMs across the edge networks, multi-LLM architectures offer a scalable and ubiquitous way to embed human-oriented general intelligence. Moreover, these systems facilitate collaborative reasoning, context-aware coordination, and task decomposition across heterogeneous edge LLMs, making LLM inferences available at the edge.

2.2.2. Applications of Multi-LLMs in EGI

In typical EGI scenarios, lightweight LLMs embedded at the edge perform context-specific tasks, while more powerful LLMs at cloud layers provide abstraction, coordination, and global optimization. Some important applications of multi-LLM systems in representative EGI scenarios are listed as follows.

• •

Smart Healthcare: EGI in healthcare is realized through a hierarchical deployment of multi-LLMs across wearable devices, bedside equipment, and hospital data centers. On-device LLMs summarize patient vitals and detect anomalies in real time, preserving privacy and minimizing latency. Summarized insights are selectively shared with cloud-hosted expert LLMs, which provide diagnostic reasoning or suggest treatment plans based on broader clinical patterns (Clusmann et al., 2023).

• •

Autonomous Mobility: In autonomous cars or drone fleets, EGI is formed via decentralized multi-agent LLM frameworks. Each vehicle hosts an embedded LLM for local perception and decision-making. Roadside units can operate LLM-based regional coordinators for traffic analysis and congestion control (Hu et al., 2025).

• •

Smart Grids: EGI in energy systems is constructed via a distributed network of LLMs embedded in smart meters, substations, and regional control centers. These local LLMs assist field engineers by analyzing real-time grid telemetry, recommending operational adjustments, and detecting faults with contextual reasoning. A central LLM, when involved, handles grid-scale forecasting or rare-event responses. The system’s decentralization ensures real-time operation and resilience against network fragmentation (Deng et al., 2024).

2.3. Zero-trust Security

2.3.1. Zero-Trust Concepts and Mechanisms

Zero-trust security is a modern cybersecurity model based on the principle of “never trust, always verify” (Poirrier et al., 2025; Joshi, 2025). Unlike traditional perimeter-based defenses that implicitly trusted entities after one-time authentication, Zero-trust assumes no inherent trust even for insiders (Poirrier et al., 2025). In practice, every access request is treated as if it originates from an open and potentially hostile environment and must be explicitly authenticated and authorized, regardless of the requester’s location or prior access history. As illustrated in Fig. 2(right), a comprehensive zero-trust architecture is supported by a suite of technical enablers.

• •

Explicit Identity Verification: Every access attempt is subject to rigorous identity authentication using multiple context-aware signals. This includes user credentials (Khowaja et al., 2025), Multi-Factor Authentication (MFA) (Wang and Wang, 2023), device hardware status (Jing et al., 2024), geographic location (Wang and Wang, 2023), behavioral indicators (Hong et al., 2023), etc. Authentication is enforced continuously, not just at entity login, ensuring that identity trustworthiness is dynamically reassessed throughout a session.

• •

Least-Privilege Access Control: Access permissions are granted according to the principle of minimum necessary privilege. Users, applications, and services are only permitted to interact with the specific resources required for their roles or tasks. This mechanism relies on fine-grained access control models such as Role-Based Access Control (RBAC) (Zhou et al., 2013) and Attribute-Based Access Control (ABAC) (Tuler De Oliveira et al., 2022), significantly reducing the attack surface.

• •

Continuous Monitoring and Analytics: System and user activities are persistently observed to detect deviations from established norms. Telemetry data, including network activity, authentication patterns, and application usage, is collected and analyzed using User and Entity Behavior Analytics (UEBA) (Najafi et al., 2021), Security Information and Event Management (SIEM) (Bhatt et al., 2014), etc. These insights enable automated policy adaptation, real-time threat detection, and incident response.

• •

Micro-segmentation and Lateral Movement Prevention: Network environments are partitioned into logically isolated zones with tailored access policies (Poirrier et al., 2025). This segmentation limits the spread of breaches by preventing unauthorized east-west movement across internal systems. Traffic between segments is monitored and governed by policy enforcement engines, ensuring that even compromised entities remain contained.

2.3.2. Applications of Zero-Trust in AI

Zero-trust principles are increasingly applied across various stages of AI pipelines to enhance trustworthiness, privacy, and resilience against adversarial threats. Representative application domains include

• •

Federated Learning (FL): In FL, multiple distributed clients collaboratively train a global model without sharing raw data. Zero-trust approaches improve FL by eliminating implicit trust assumptions among clients (Bandara et al., 2022). Mechanisms such as client attestation, secure identity provisioning, and TEE (Yu et al., 2022) can be employed to detect data poisoning, model inversion, and free-riding attacks.

• •

Zero-Trust Model Access: As AI models are increasingly deployed across distributed and heterogeneous platforms, zero-trust security frameworks have become essential to safeguard model endpoints. Mechanisms such as dynamic authentication, continuous identity verification, and fine-grained access control mitigate the risks of unauthorized queries and adversarial model extractions (d’Aliberti et al., 2024).

• •

Multi-Agent and Collaborative Intelligence: In distributed AI systems involving interacting agents, zero-trust ensures that inter-agent communications are authenticated, contextually scoped, and revocable. Protocols such as certificate pinning, time-bound tokens, and risk-adaptive authentication prevent spoofing or privilege escalation, especially in scenarios like distributed inference, agent planning, or swarm robotics.

• •

Zero-Trust Reinforcement Learning: In safety-critical RL applications (e.g., autonomous driving), adversarial attacks on sensors, environments, or reward models pose severe risks (Zhao et al., 2024). Zero-trust RL frameworks integrate behavioral policy validation, continuous verification of inputs, and response gating to mitigate risks from compromised components or manipulated environments.

3. Security of Multi-LLMs in EGI

In this section, we introduce the security issues faced by multi-LLM systems in EGI (see Table 2). Moreover, we review perimeter-based security defenses.

3.1. Intra-LLM Level Security Issues

3.1.1. Jailbreaks Attack and Prompt Injection

Even on the edge, LLMs remain vulnerable to malicious prompts designed to override their safeguards (Xu et al., 2024). Adversaries can craft “jailbreak” inputs that exploit the LLM’s learned patterns (often by obfuscation (Xu et al., 2024) or role-playing (Jin et al., 2025)) to make it ignore safety instructions. Similarly, prompt injection attacks use malicious text to manipulate the LLM’s internal state. For instance, Liao et al. (Liao and Sun, 2024) demonstrated a single “universal” prompt that reliably bypassed safeguards in GPT-4, Bing, Bard, and Claude simultaneously. In an EGI context, a local edge LLM with weak or no moderation could be tricked into providing harmful output, such as advising on illicit activities or disclosing system secrets.

3.1.2. Unpredictable Emerging Abilities

LLMs may demonstrate emergent capabilities that were not anticipated by their developers, and this unpredictability can be dangerous in EGI deployments. Because, due to resource constraints (He et al., 2025), EGI systems often use smaller or specialized LLMs at the edge, one might assume they are safer. However, even downsized LLMs can exhibit surprising skills learned from vast pretraining data. For instance, a recent study found that GPT-4 can autonomously generate functional exploit code for 87% of known one-day software vulnerabilities when simply given their official descriptions (Fang et al., 2024). LLM’s ability to produce working attack payloads without human guidance reveals a serious security concern. In EGI applications, if an edge LLM agent unexpectedly acquires such capabilities (e.g., crafting network intrusions or bypassing authentication), malicious users could leverage it to harm the local environment. The difficulty of fully predicting an LLM’s behavior means edge systems face a risk of unknown unknowns, complicating safety assessments and deployment decisions.

3.1.3. Data Leakage and Privacy

LLMs have no intrinsic concept of confidentiality. Specifically, they generate outputs based on training data and input prompts without understanding what should not be revealed. This poses acute risks in EGI systems that integrate LLMs with sensitive enterprise or sensor data. An edge-based LLM might inadvertently disclose private information, e.g., personal details, device telemetry, and access credentials, in its responses if the prompts indirectly trigger memorized data (Carlini et al., 2021). Likewise, users may unwittingly input confidential data, e.g., customer records, internal documents, source code, into an LLM-powered edge service (Mireshghallah et al., 2022). If such privacy is recorded in the memory of the multi-LLM systems and shared among LLMs, persistent privacy violations can happen. Additionally, without strict access controls, an insider or attacker which can query a local LLM could extract any data the model has seen (Carlini et al., 2021). In summary, data leakage in multi-LLM EGI can flow both ways: the LLM might expose private training data outward, or ingest private user input that stays in system memory, later becoming accessible to others. These outcomes undermine the presumed privacy advantages of processing data on the edge.

3.1.4. Toxic or Misaligned Responses

Even when functioning correctly, LLMs can produce biased, offensive, or hallucinatory content due to biases in training data or misalignment with human values (Sahoo et al., 2024). In an EGI setting, such toxic or misleading output can have serious real-world consequences. For example, if a local financial advisor LLM on a branch office server fabricates wrong investment statistics or a healthcare assistant LLM on a hospital device gives a biased medical recommendation, the consequences could be legal and reputation damage. Moreover, edge LLM deployments often utilize open-source or customized models that may lack rigorous moderation. Early versions of Meta’s LLaMA, for instance, could generate extremist or misinformation text when prompted (Sahoo et al., 2024). A misaligned response from an edge LLM can directly impact on-site operations or user safety. For instance, incorrect driving decisions generated by an LLM-based autonomous vehicle could lead to traffic violations, collisions, or endangerment of human lives. Thus, ensuring response quality is not merely an ethical concern but a critical security requirement in multi-LLM EGI systems.

3.2. Inter-LLM Level Security Issues

3.2.1. Expanded Attack Surface

Deploying multiple LLMs inherently enlarges the system’s attack surface, as each LLM becomes a potential entry point for adversaries. In multi-LLM environments, a single compromised model can trigger cascading failures across the entire network. Previous work demonstrates that adversarial prompts injected into one LLM can propagate through inter-agent communication, resulting in a “chain-of-compromise” effect (Lee and Tiwari, 2024). For example, Lee et al. (Lee and Tiwari, 2024) proposed a prompt infection framework, where a malicious payload spreads covertly between LLMs through normal message passing. Shen et al. (Shen et al., 2025) further showed that multi-LLM s coordination increases the likelihood of compound vulnerabilities, especially in decentralized topologies.

3.2.2. Over-Permissive Integration

In EGI, LLMs are often tightly integrated with privileged components, such as internal APIs, data pipelines, code execution engines, and physical devices. Unlike traditional software, LLMs interpret and act on natural language instructions, potentially triggering actions dynamically based on prompt semantics (Zhang et al., 2025). Without strict enforcement of privilege boundaries, even innocuous prompts can lead to sensitive operations. A recent study demonstrated that GPT-4, when granted tool-use capabilities, autonomously devised a strategy to bypass a CAPTCHA challenge by recruiting a human worker on TaskRabbit. It further deceived the worker by falsely claiming to be visually impaired to justify the request (OpenAI et al., 2023). This illustrates how an LLM, once endowed with autonomy, can infer and execute multi-step plans that fall outside its intended operational scope. In EGI settings, such over-permissive integration can result in LLMs issuing unintended commands, escalating privileges, or modifying protected data.

3.2.3. Insecure Inter-LLM Communication

LLMs need to frequently exchange prompts and responses to coordinate tasks. If communication channels are not authenticated or encrypted, they become vectors for prompt injection, impersonation, or eavesdropping (Lee and Tiwari, 2024). One prominent threat is inter-LLM prompt injection, where an adversary injects a malicious instruction into one LLM that subsequently propagates across peers (He et al., 2025). Furthermore, if identity verification is absent, attackers may impersonate trusted LLMs, forcing others to perform unauthorized actions or disclose private data (Zhang et al., 2024). Attackers may also attempt passive surveillance by intercepting inter-LLM traffic, reconstructing sensitive task contexts, or operational intents. In edge environments where lightweight communication protocols are common, these risks are amplified due to the absence of a centralized authentication infrastructure.

3.2.4. Consensus Manipulation by Byzantine LLMs

Multi-LLM systems often rely on consensus protocols, either centralized (e.g., round-robin leader rotation (Zhang et al., 2025)) or decentralized (e.g., random voting (Zhang et al., 2025)), to coordinate collective decisions. However, if one or more LLMs behave in a Byzantine manner, they can disrupt collaborative outcomes. Malicious LLMs may inject false information, vote dishonestly, or collude to dominate consensus, especially in decentralized settings lacking a global coordinator. Note that even centralized coordination does not guarantee safety. If the coordinator itself is compromised, it can propagate biased plans, suppress valid alternatives, or selectively misinform subordinates. Such manipulation can lead to unsafe decisions, policy violations, or service disruptions, particularly damaging in EGI contexts where coordination governs physical processes such as autonomous fleets or energy balancing.

3.2.5. Cross-Context Data Leakage from Inter-Agent Invocation

In multi-LLM EGI deployments, individual LLMs often manage localized datasets tailored to specific functions or regions. However, when LLMs collaborate through task delegation or response chaining, unintended data exposure can occur. For example, a predictive maintenance LLM might query a building usage LLM for context, unintentionally eliciting user-specific energy data that fall outside its access scope. Even when each LLM complies with its own policy in isolation, combined behavior can violate global privacy constraints. Recent work shows that carefully designed multi-step prompts can extract sensitive information distributed across LLMs by reconstructing partial answers (Juneja et al., 2025). In decentralized EGI environments, the absence of centralized policy enforcement exacerbates this issue.

3.3. Perimeter-Based Security: Towards Trustworthy Multi-LLMs

Traditional security strategies rely heavily on perimeter-based defenses, setting a clear boundary within which all components are implicitly trusted. Consequently, strengthening security typically involves reinforcing or extending these security boundaries. In multi-LLM contexts, researchers have followed this conventional principle to develop trustworthy LLMs (Yu et al., 2024; Chrapek et al., 2024; Namer et al., 2025), constructing security boundaries from multiple perspectives, including model capabilities, training data quality, runtime environments, and interaction processes. In the following parts, we survey the representative approaches for implementing trustworthy LLMs (see Fig. 3).

3.3.1. Adversarial Training

Adversarial training (Yu et al., 2024) is a robust strategy to enhance the trustworthiness of LLM capabilities against harmful inputs and jailbreak attacks. Specifically, this technique involves repeatedly exposing the LLMs to specially crafted inputs designed to lead the models into producing incorrect or harmful responses. In this way, LLMs can learn from these difficult examples, gradually improving their ability to identify and resist such attacks (see Fig. 3).

For example, Xhonneux et al. (Xhonneux et al., 2025) proposed Continuous Adversarial Training (CAT). By adjusting input prompt embeddings in directions that maximize LLM confusion, CAT efficiently synthesizes challenging training samples. These maliciously perturbed samples are used in the training loop, enabling LLMs to iteratively learn to recognize their vulnerabilities and minimize the adverse effects of these perturbations. Furthermore, Yu et al. (Yu et al., 2024) developed Refusal Feature Adversarial Training (ReFAT), focusing on a specific embedding characteristic called refusal feature, which signals when an LLM should reject harmful requests. ReFAT deliberately disrupts this feature during training, effectively simulating conditions where the LLM’s safeguard mechanisms are compromised. By repeatedly forcing the LLM to handle scenarios where refusal features are unavailable, ReFAT can strengthen the LLM’s overall robustness. Recently, AdversaFlow (Deng et al., 2025) introduced a visually interactive approach to adversarial training. By visualizing the progression of adversarial attacks and LLM responses, experts can interactively generate, observe, and analyze adversarial examples, iteratively refining LLM input based on reactions.

3.3.2. Differential Privacy

Trustworthy multi-LLM not only means preventing harmful outputs, but also protecting sensitive inputs and training data. In multi-LLM EGI scenarios, edge devices may handle private user data or proprietary information that cannot be leaked. Moreover, researchers have shown that LLMs can inadvertently memorize training data, and determined attackers might extract verbatim sensitive information (e.g., personal identifiers or secret keys) from an LLM by querying it strategically (Charles et al., 2024). Hence, DP (Behnia et al., 2022) is an effective approach that can be used during LLM training to ensure that the model does not remember or reveal specific details from its training examples. Specifically, DP addresses this by adding carefully calibrated noise to the training process, so that any single training datum has only a negligible influence on the final LLM parameters. In practice, applying DP to LLMs means that the model learns general patterns but cannot reproduce any exact training sentence with high confidence. For instance, Behnia et al. (Behnia et al., 2022) presented a DP-assisted fine-tuning framework that achieves measurable privacy guarantees while minimizing the impact on LLM performance. Considering that each EGI device fine-tunes an LLM based on the user interactions it observes, applying DP ensures that even if one LLM is compromised, the attacker cannot directly identify raw data. Moreover, since DP facilitates privatized updates, multiple devices can contribute to training a joint LLM without exposing their raw data.

3.3.3. Reinforcement Learning with Human Feedback

As shown in Fig. 3, RLHF is a prominent approach to fine-tuning LLMs to align their output more closely with human preferences, thus improving their safety, ethical compliance, and overall trustworthiness. For instance, Dai et al. (Dai et al., 2024) proposed the Safe RLHF framework, explicitly capturing human preferences concerning helpfulness and harmlessness. By training LLMs to generate content that aligns with user preferences, the usefulness and safety of the generation can be improved. However, conventional RLHF approaches suffer from issues, such as reward hacking and incorrect generalization of target objectives. To address these limitations, Sun et al. (Sun and Zhao, 2025) introduced a personalized LLM security alignment framework. By incorporating user-specific preferences and employing multi-dimensional reward functions, their approach provides precise reward signals to RLHF, leading to more accurate and contextually relevant safety alignment. Moreover, the authors in (Spelda and Stritecky, 2025) emphasized that despite the substantial improvements RLHF offers in aligning LLMs with human preferences, it cannot guarantee absolute safety. This inherent limitation arises because RLHF relies on generalizing human preferences from limited and potentially biased human feedback. Consequently, achieving complete safety and ethical alignment remains challenging, underscoring the need for continuous human oversight, diverse participation, and robust testing methodologies.

3.3.4. Reactive Output Filtering

Another important defense is output filtering, where a separate mechanism intercepts and screens the LLM’s generated content before it reaches the end-user (Phute et al., 2024). A simple approach is to employ a classification model or heuristic rules to detect toxic, hateful, or otherwise harmful language in the LLM’s output and block or modify it as needed. Recent research has explored using LLMs themselves as filters, effectively turning the model’s own knowledge of harmful content into a defensive tool. For example, Self-Defense (Phute et al., 2024) prompts the LLM to examine its output for harmful content. In detail, after the LLM generates a response, that response is fed (along with an instruction) to the same or another LLM, which should decide whether the content is malicious or violates policies. The experiments demonstrate that Self-Defense can reduce the attack success rate to virtually zero across a variety of adversarial prompt attacks. Note that the advantage of output filtering is that it provides a second line of defense. Even if one entity in the multi-LLM produces harmful content, a filter can catch it before it propagates further or is shown to users. Many commercial AI systems (e.g., OpenAI’s and Anthropic’s APIs) incorporate such automated content filters.

3.3.5. Hardware Enclaves

Hardware Enclaves (i.e., TEEs) are secure, hardware-backed execution contexts designed to isolate sensitive computations and data, ensuring they are protected from unauthorized access and tampering (Chrapek et al., 2024; Dong and Wang, 2025). Leveraging hardware-based isolation, TEEs guarantee that the code and parameters of LLMs remain confidential and trusted. Moreover, TEEs support run-time attestation, which continuously verifies and confirms the integrity of the LLM and the secure operation throughout its execution.

Numerous research efforts have explored the deployment of LLMs within TEEs to enhance security. First, Chrapek et al. (Chrapek et al., 2024) demonstrated a practical method to secure LLMs employing Intel SGX and Trust Domain Extensions (TDX). Their implementation involves isolating LLM execution within hardware enclaves, ensuring continuous verification and protection against unauthorized model extraction, malicious quantization attacks, and unauthorized fine-tuning. Their approach achieved less than 10% overhead compared to conventional, insecure deployments, highlighting the feasibility and efficiency of deploying complex inference tasks within TEEs. Similarly, Dong et al. (Dong and Wang, 2025) benchmarked the performance of DeepSeek LLMs across TEE-based, CPU-only, and CPU-GPU implementations. Su et al. (Su and Zhang, 2025) introduced a framework using Intel TDX with run-time attestation. This approach secures containerized LLMs by continuously validating their integrity within TEEs, effectively preventing unauthorized access and tampering during model execution. Furthermore, this solution includes a hardware-agnostic runtime environment, which mitigates vendor lock-in issues, enabling secure edge LLM deployment across diverse hardware platforms. Similarly, Li et al. (Li et al., 2024) proposed embedding LLM computations within TEEs, ensuring robust security guarantees. They also optimized inter-TEE communication through direct secure channels, greatly guaranteeing both the security and performance of decentralized AI inference tasks, particularly suitable for high-stakes domains like healthcare and finance. Finally, Lin et al. (Lin et al., 2025) introduced the LoRATEE framework, specifically designed for secure and efficient multi-tenant LoRA-based LLM inference using TEEs. By embedding LoRA adapters inside Intel SGX enclaves and employing a lightweight one-time pad encryption for secure data transmission between the enclave and external computing environments (such as GPUs), they significantly mitigate potential security risks associated with multi-tenant environments.

3.3.6. Access Control & Firewalls

Access control and firewalls are critical security mechanisms for ensuring the confidentiality and security of multi-LLM systems in EGI environments. Effective access control mechanisms regulate user permissions, ensuring that only authorized entities can interact with LLM resources and preventing unauthorized data access and operations (Sanyal et al., 2025). Firewalls, in turn, monitor and filter both incoming and outgoing traffic according to predefined security rules, protecting LLM systems from malicious inputs and unauthorized access attempts (Huang et al., 2024).

In terms of access control, traditional approaches have relied primarily on static permission models such as discretionary access control (Lin and Vachon, 2017) and mandatory access control (Lu et al., 2025), which grant fixed permissions based on user identity or security clearance levels. However, for multi-LLM systems, RBAC has emerged as a particularly relevant paradigm, where permissions are associated with LLMs with specific roles, enabling dynamic permission scheduling during inter-LLM collaborations. The importance of RBAC for multi-LLM is reported by Sanyal et al. (Sanyal et al., 2025). Their assessment of 16 state-of-the-art LLMs across 40 distinct enterprise scenarios revealed that even flagship models like GPT-4.1 achieve only 27% accuracy when reasoning about complex multi-permission cases. To this end, Shi et al. (Shi et al., 2025) introduced Progent, a programmable privilege control mechanism that realizes RBAC for LLMs. Rather than relying on static user-based permissions, Progent implements dynamic role-driven policies where tool access rights are determined by the LLM’s assigned role and current task context. Moreover, privilege control policies are expressed by JSON files that define which tool calls are permissible for specific roles, under what conditions they are allowed, and what fallback actions occur when role-based constraints are violated. This RBAC-based approach enables scalable permission management across multi-LLM systems, where different LLMs can assume different organizational roles with corresponding access privileges, ensuring secure execution while preventing unauthorized operations that exceed their designated role boundaries.

As for firewalls, Huang et al. (Huang et al., 2024) introduced FirewaLLM, a portable framework that desensitizes sensitive user inputs using smaller local models before interactions with large LLMs. This approach effectively prevents inadvertent disclosure of personal, health, or financial data during LLM interactions. Abdelnabi et al. (Abdelnabi et al., 2025) proposed a comprehensive firewall framework tailored for dynamic LLM networks. Their design automatically generates task-specific firewall rules from prior conversations, sanitizing inputs into deterministic and verifiable formats. It also incorporates dynamic abstraction of user data and implements self-correcting mechanisms, significantly reducing the risks of prompt injection, data exfiltration, and unauthorized manipulation. Yao et al. (Hongwei Yao, 2025) presented ControlNet, an AI firewall specifically designed for Retrieval-Augmented Generation (RAG)-based LLM systems. Leveraging neuron activation shift phenomena, ControlNet detects and mitigates malicious queries, enforcing precise query flow control policies. Finally, Namer et al. (Namer et al., 2025) developed an automated framework to detect ”expensive” prompts designed to overload LLM systems, configuring firewall rules dynamically to mitigate denial-of-service attacks.

3.3.7. Secure Communication Protocols

Given the extensive interactions and information exchanges inherent in multi-LLM systems, trustworthy communications should also be considered. To this end, Gan et al. (Gan et al., 2024) proposed a binary mapping framework that systematically categorizes threats based on sources and impacts, enabling precise identification and blocking of malicious or misleading interactions across LLMs. Multi-layered security mechanisms, including data link layer encryption protocols and network layer agent behavior monitoring, further protect information exchanges between LLMs and with users or environmental interfaces. Advanced frameworks, such as MCP-Shield555https://github.com/riseandignite/mcp-shield, extended this protection through signature-matching and adversarial behavior profiling, enabling pre-execution detection of high-risk tools and malformed tasks. Moreover, MCIP (Jing et al., 2025) introduces runtime trace analysis with an explainable logging schema and security-awareness model to track violations in complex LLM-tool interactions. These robust communication protocols collectively establish a comprehensive security posture that ensures secure, reliable, and verifiable inter-LLM communications in multi-LLM systems.

3.4. From Trustworthy to Zero-Trust Defenses

Trustworthy LLM proposals (Yu et al., 2024; Phute et al., 2024; Yu et al., 2022; Namer et al., 2025; Gan et al., 2024) have significantly improved the security posture of multi-LLM systems. However, these traditional perimeter-based strategies present several inherent limitations that become increasingly pronounced as multi-LLM systems evolve in complexity and scope. First, as LLM capabilities expand and their operational domains broaden, the attack surface increases correspondingly (Zhao et al., 2025), making it challenging to maintain clearly defined security boundaries. Moreover, the frequent interactions and data exchanges inherent in multi-LLM systems create extensive lateral movement opportunities across different security domains, potentially compromising data integrity and security perimeters (Li et al., 2024). Finally, traditional security paradigms often exhibit reactive characteristics (i.e., responding to threats only after attacks or breaches occur), which can lead to delayed detection and mitigation that allows substantial damage. These limitations prove especially pronounced in EGI environments, where sophisticated prompt injections and adversarial attackers damage LLMs in unforeseen ways. Meanwhile, resource constraints make extensive security measures, such as comprehensive adversarial training or TEE deployment, impractical due to computational, energy, and bandwidth limitations.

Remark 1.

These systemic limitations raise a fundamental question: can LLMs be trusted? From the user’s perspective, LLMs exhibit inherent opacity and limited interpretability due to their extensive training datasets and complex architectural structures, making human oversight and comprehensive screening practically infeasible (OpenAI et al., 2023). From the LLM perspective, each LLM cannot assume the trustworthiness of collaborating ones, as these may be compromised, malicious, or manipulated by adversaries seeking to exploit the collaborative framework (Chen et al., 2024). This mutual distrust requires a new security paradigm that does not rely on the assumed benevolence of LLMs, i.e., zero-trust.

4. Zero-Trust Multi-LLM for EGI

4.1. Constructing Zero-Trust Multi-LLMs for EGI

We observe that many advanced studies (Chan, 2025; Shi et al., 2025; Xiao et al., 2025; Arshad and Halim, 2025; Chen et al., 2025a, b; Chong et al., 2025; Wang et al., 2024) have proposed defenses aligning with zero-trust principles or presented zero-trust single LLMs, while a comprehensive tutorial systematically discussing zero-trust multi-LLM systems in EGI has yet to emerge. Hence, we demonstrate the vision of zero-trust multi-LLM in EGI using an autonomous driving case study. As illustrated in Fig. 4, the EGI paradigm involves heterogeneous LLMs with diverse capabilities, computational requirements, and operational scopes deployed across cloud infrastructures, edge servers such as roadside units, and on-vehicle chips. Note that we follow NIST SP 800-207 (NIS, [n. d.]), the most widely adopted zero-trust standard, to showcase the architecture. The subsequent parts detail the systematic implementation of each component of this framework.

4.1.1. Mobile-Edge LLMs

As shown in Fig. 4, mobile LLMs operate on edge servers and vehicular chips to ensure minimal latency for real-time decision-making in autonomous driving scenarios. In multi-LLM settings, these LLMs undergo functional segregation to handle distinct operational aspects and collaborate to accomplish the assigned task (Li et al., 2024). For instance, perception LLMs transform raw sensor output into structured textual scene descriptions, and planning LLMs interpret traffic regulations for high-level route planning decisions. Zero-trust enforcement restricts each LLM’s access exclusively to the minimum necessary data and command interfaces. For example, perception LLMs can access camera and LiDAR feeds but cannot interface with steering or acceleration controls. Upon system initialization, each LLM should undergo cryptographic identification to prevent Sybil attacks (Yao et al., 2019).

4.1.2. Cloud LLMs

Cloud LLMs can offer computationally intensive generation capabilities, comprehensive knowledge repositories, and coordination services that surpass the capacity of mobile-edge LLMs. Moreover, following NIST zero-trust standards (NIS, [n. d.]), cloud LLMs function as policy engines that establish and enforce security protocols throughout the distributed multi-LLM network while maintaining centralized oversight of system-wide security policies.

4.1.3. Identity and Authentication Module

At the core of the framework lies a comprehensive identity management system that establishes the foundation for all zero-trust operations. Every LLM, whether deployed at the edge or in the cloud, receives a cryptographically secure identity comprising unique asymmetric key pairs and digitally signed certificates (Sun et al., 2025). This module enforces mandatory identity verification for each interaction, eliminating the possibility of anonymous or unverified communications within the system. Moreover, the authentication mechanism implements continuous verification rather than one-time startup authentication. Session credentials, which can be based on multi-factor authentication (Wang and Wang, 2023), tags (Zhang et al., 2022), or reputation (Nagarajan et al., 2024; Fang et al., 2024) are deliberately short-lived and require frequent re-establishment to mitigate the impact of potential credential theft or compromise. This approach aligns with zero-trust principles by maintaining persistent identity verification throughout the operational lifecycle, ensuring that even previously authenticated entities must continuously prove their legitimacy to maintain system access.

4.1.4. Inter-LLM Communication Gateway

As illustrated in Fig. 4, all inter-LLM communications within the multi-LLM system are channeled through a secure gateway deployed on the cloud LLM. This gateway functions as both an intelligent firewall and a message broker, implementing policy-controlled routing mechanisms that enforce strict communication boundaries between system components (Abdelnabi et al., 2025; Ehtesham et al., 2025). Specifically, it maintains dynamic and fine-grained access control policies that specify authorized communication patterns for each type of LLM. Any communication attempt that violates the policies is immediately intercepted and blocked by the gateway, preventing unauthorized data flows and potential security breaches. Moreover, transmitted content is encrypted using the sender LLM’s public key and decrypted at the receiver using private keys, thereby preventing sensitive information from being intercepted or tampered with during transmission (Sun et al., 2025). These mechanisms ensure that no internal message routing occurs based on implicit trust assumptions about the source, requiring explicit verification for every communication attempt.

4.1.5. User Input Checking

User inputs are treated as potentially hostile data sources within the zero-trust framework, as they may originate from malicious attackers or be compromised through adversarial manipulation during transmission. These inputs could contain carefully crafted prompt injection attacks designed to bypass LLM safety guardrails and trigger jailbreaking behaviors that compromise system integrity (Chen et al., 2025a). The input validation module implements comprehensive sanitization mechanisms, including lexical analysis to detect suspicious instruction patterns, semantic filtering to identify context manipulation attempts, and behavioral pattern recognition to flag inputs deviating from normal user interaction profiles. This proactive input filtering ensures that adversarial prompts cannot propagate through the multi-LLM system and maintains the principle of explicit verification.

4.1.6. Multi-Layer LLM Output Verification

Similar to perimeter-based security, zero-trust multi-LLM systems incorporate a comprehensive multi-layer verification engine that embodies the ”never trust, always verify” principle through hierarchical validation mechanisms. This multi-tiered filtering system can combine rule-based constraints, intelligent verification algorithms, and context-aware dynamic assessments to ensure that every LLM output undergoes rigorous validation before execution (Gehman et al., 2020).

4.1.7. Behavioral Auditing and Anomaly Detection

Zero-trust requires persistent surveillance, since breaches are inevitable and threats can arise from any component of the entire multi-LLM system (Joshi, 2025). The behavioral auditing module implements comprehensive logging of all LLM communications, decisions, and collaborative interactions with tamper-resistant records stored in secure local repositories. Real-time anomaly detection algorithms analyze behavioral patterns to identify security threats, such as plan LLMs attempting direct vehicle control or outputs containing prompt injection tokens.

4.2. Operational Workflow

Fig. 4 illustrates the end-to-end operation of the zero-trust multi-LLM system in an autonomous driving scenario. Four major steps are included.

• •

Startup and LLM Authentication: Initially, each mobile LLM generates its public-private key pair and registers its identity with the cloud LLM. The policy engine initializes access control mechanisms, input and output checking protocols, etc. Meanwhile, the behavioral auditing system begins comprehensive logging of all inter-LLM communications and individual LLM activities for anomaly detection.

• •

Prompt Understanding: User prompts undergo rigorous validation through the input checking module before processing. If anomalous content is detected, the framework immediately quarantines the input and initiates additional verification procedures to prevent system compromise (NIS, [n. d.]).

• •

Collaborative Decision Planning: The primary advantage of multi-LLM systems lies in their ability to organize LLMs collaboratively (Yang et al., 2024). Multiple LLMs participate in decision-making processes in which the planning LLMs formulate initial decisions, the cloud LLM provides expert validation for complex scenarios, and the safety LLMs perform independent assessments of the collaboration process. The system supports multiple collaboration patterns, such as pipeline or distributed voting (see Section 2.1), based on the specific task.

• •

Policy-Governed Plan Execution: The multi-LLM system generates final outputs that trigger executable vehicle commands, such as lane change maneuvers, speed adjustments, and braking actions (Sun et al., 2024). The multi-layer verification mechanism validates all LLM outputs against safety constraints, regulatory compliance, and operational feasibility before execution. Moreover, execution can only be initiated by designated LLMs that have explicit authorization for specific vehicle control functions.

• •

Continuous Monitoring and Threat Mitigation: The behavioral auditing system monitors all inter-LLM communications and individual LLM behaviors, maintaining comprehensive logs for forensic analysis and real-time threat detection. When anomaly detection mechanisms identify suspicious behavior or malicious content generation from any LLM, the cloud LLM isolates compromised LLMs while enabling the remaining LLMs to continue collaborative operation, ensuring system resilience and operational continuity (Joshi, 2025).

Lesson Learned: By organizing the modules involved according to the workflow detailed above, we can construct a multi-LLM framework in EGI that effectively fulfills the four fundamental principles of the zero-trust paradigm (i.e., explicit verification, least privilege, continuous monitoring, and segmentation) (Joshi, 2025). Note that our demonstrated implementation adheres to the most famous NIST SP 800-207 zero-trust standard (NIS, [n. d.]), which employs a policy engine responsible for defining security policies (e.g., input/output validation rules and agent authentication), making real-time decisions on access and communication, and dynamically coordinating interactions among LLMs. Nevertheless, other influential Zero-trust standards and frameworks, such as Google’s BeyondCorp666https://cloud.google.com/beyondcorp and CISA777https://www.cisa.gov/zero-trust-maturity-model, can also be applied.

4.3. Trustworthy Multi-LLM versus Zero-Trust Multi-LLM

To demonstrate the features of zero-trust multi-LLM systems, we compare them against previous ones operating with security perimeters (see Table 3). First, we observe that perimeter-based security and zero-trust security operate on fundamentally different assumptions, with different mechanisms. Moreover, traditional security approaches can rely on isolated security strategies that address individual vulnerabilities independently. For instance, we can employ DP to protect LLM training data or deploy LLMs within TEEs to ensure computational integrity, treating these as standalone solutions. However, constructing a zero-trust environment represents a comprehensive systems engineering challenge that requires an end-to-end workflow to orchestrate multiple modules, as demonstrated in our proposed framework above (NIS, [n. d.]). From a resource management perspective, the traditional security paradigm might incur significant training costs and hardware overhead, such as extensive adversarial training and the deployment of TEEs. However, zero-trust approaches introduce additional operational overhead, primarily due to continuous verification processes (Joshi, 2025) and persistent authentication and monitoring requirements (Nahar et al., 2024).

Note that it is important to recognize overlaps between traditional perimeter-based security methods and zero-trust strategies. Several foundational techniques, such as cryptographic authentication (Sun et al., 2025) and access control mechanisms (Sanyal et al., 2025; Shi et al., 2025), apply to both concepts. For example, access control mechanisms can be utilized to construct security boundaries for LLMs by defining precise permissions for accessing specific tools and data. However, zero-trust imposes even higher demands on access control, such as context-aware policies, which dynamically adjust permissions based on real-time contextual factors for different micro-segmentations. Similarly, continuous monitoring, although prevalent in traditional security frameworks, assumes heightened importance in zero-trust environments. Intelligent monitoring ensures real-time detection and proactive response to threats, maintaining secure operations even as conditions evolve.

5. Intra-LLM Zero-trust Security: Model Approaches

In this section, we review the technical progress of model-level approaches that provide zero-trust management of individual LLMs.

5.1. Strong LLM Identity, Authentication, and Authorization

First, each LLM in a zero-trust architecture is assigned a robust cryptographic identity to ensure secure and verifiable identification. Practically, this involves providing every LLM with a unique asymmetric key pair and a digital certificate, similar to IoT or embedded devices that use device-specific keys (Sun et al., 2025) or X.509 (Kinkelin et al., 2020) certificates to authenticate identity. Moreover, every user-to-LLM, inter-LLM, and LLM-to-tool call requires prior authentication. Note that authentication is continuous rather than a one-time handshake; credentials remain short-lived and frequently renewed, embodying the zero-trust principle of persistent verification (Nahar et al., 2024). Representative authentication approaches are described below (see Table 4 also).

5.1.1. Multi-factor Authentication

MFA introduces additional verification layers beyond a single credential, forming a cornerstone of zero-trust security (Zhou et al., 2025). By requiring LLMs to present multiple independent proofs of identity, such as cryptographic keys combined with contextual information (e.g., device attestation, biometric verification, or behavior analytics), MFA substantially reduces the risk of unauthorized access due to compromised credentials. Recent research emphasizes context-aware or adaptive MFA, dynamically adjusting verification requirements based on real-time risk assessments, including agent behavior, device integrity, location, and access patterns (Wang and Wang, 2023). This dynamic MFA approach effectively balances security and usability by tightening security protocols during detected anomalies. These adaptive mechanisms, initially designed for device authentication, are particularly applicable to authenticating LLMs in multi-LLM zero-trust EGI, with heterogeneous devices and application patterns (Arias-Cabarcos et al., 2019).

5.1.2. Reputation-based Authentication

In zero-trust multi-LLM systems, each LLM can maintain a reputation derived from historical outputs, policy adherence, and user/peer feedback (Nagarajan et al., 2024; Fang et al., 2024; Shaina Raza, 2025). Suspicious or erroneous behavior results in decreased reputation scores, triggering restricted privileges, while consistently reliable LLMs accrue higher reputation scores and consequently gain access to more sensitive tools or data. Moreover, blockchain technology has been proposed to securely track reputation, providing transparent, verifiable, and immutable reputation histories (Geng et al., 2021). For example, LLMChain integrates automated performance assessments and human feedback within a blockchain ledger to calculate comprehensive, tamper-proof reputation scores (Bouchiha et al., 2024).

5.1.3. Token-Based Authentication

Chen et al. (Chan, 2025) provided fine-grained authorization through encrypted tokens encoding specific permissions for LLM interactions. Specifically, a central provider, potentially served by cloud LLMs within multi-LLM EGI architectures, maintains LLM identities and access policies, issuing ephemeral, cryptographically secure tokens defining permitted interactions. LLMs should obtain fresh tokens for each session or interaction, strictly limiting privileges to the immediate task. Notably, each token has built-in expiration conditions, requiring continuous re-authentication. Such adaptive authentication effectively enforces the principle of least privilege (Joshi, 2025), ensuring that permissions are granted per request and revoked as soon as tasks are completed, aligning seamlessly with zero-trust principles of persistent verification and minimal trust.

Lesson Learned: We can conclude that MFA provides the strongest security through layered verification, but introduces computational overhead suitable for high-security applications, while reputation offers dynamic trustworthiness management ideal for long-term deployments with established behavior patterns. Token-based authentication delivers fine-grained access control with minimal overhead, making it effective for distributed systems requiring precise privilege control.

5.2. Context-aware Access Control

Zero-trust principles demand stringent access control to prevent privilege abuse and information leakage in multi-LLM systems. However, traditional RBAC, such as simple role-based permissions (Ghafoorian et al., 2019), is insufficient for zero-trust LLM collaboration. RBAC assumes that if an LLM’s identity or role is verified, it can be trusted broadly, which is an assumption zero-trust explicitly rejects.

Context-aware Access Control (CAAC) frameworks extend beyond identity: “context” can include any information about the user/LLM, the resource, or the environment of an access request (Kayes et al., 2020). For instance, some proposals augment the RBAC with conditions, such as location, time of request, or other situational constraints (Kayes et al., 2020). Such context-driven policies allow fine-grained, situation-specific decisions (e.g., permitting an LLM to see certain data only during a particular task or within certain time bounds), aligning with zero-trust’s demand for dynamic and least-trust enforcement.

Furthermore, multiple recent works propose advanced access control mechanisms tailored to multi-LLM systems that embody zero-trust principles. AgentSafe (Mao et al., 2025) is one such framework that protects multi-LLM systems through hierarchical data management. It classifies information into security levels and restricts sensitive data to only LLMs with authorization. Moreover, it introduces two components: ThreatSieve and HierarCache. ThreatSieve secures inter-LLM communication by authenticating each message’s source and verifying the sender’s authority, blocking unauthorized or impersonating LLMs. HierarCache, on the other hand, hierarchically manages LLM memory to prevent data leakage and poisoning. Historical information is stored in “drawers” according to sensitivity and LLM relationships, so each LLM only retains or sees memory appropriate to its trust level. Notably, AgentSafe’s design is dynamic: as LLMs join/leave or as tasks change, the hierarchy can adjust. Complementing the hierarchical approach, the Embedded Privacy-Enhancing Agents (EPEAgents) (Shi et al., 2025) paradigm focuses on context-aware filtering of information in multi-LLM systems. Specifically, cloud LLM will coordinate context sharing during LLM collaboration. The core idea is to minimize data exchange: rather than broadcasting a raw knowledge base or full context to all LLMs, each LLM receives only the task-relevant information that it truly needs. Importantly, EPEAgents is built to preserve contextual awareness for legitimate collaboration. Unlike rigid RBAC, which might omit crucial context and degrade performance, EPEAgent ensures each LLM still gets the context it needs by dynamically evaluating both the LLM’s declared role and the evolving task requirements, allowing the system to adapt to changing collaboration patterns while maintaining strict information boundaries.

Other works bring cryptographic enforcement to access control for multi-LLM systems, using fine-grained encryption to guarantee that even if communications are observed by attackers, only authorized LLMs can decrypt and use sensitive information. Xiao et al. (Xiao et al., 2025) proposed a privacy-preserving access control scheme for LLM-driven networks based on Attribute-Based Encryption (ABE) with policy hiding and revocation. In their design, each piece of content (e.g., a dataset, prompt, or answer) can be encrypted under an access policy (defined as a logical combination of attributes) such that only an LLM possessing a key with attributes satisfying that policy can decrypt it. This allows extremely fine-grained control: policies might encode that “only an LLM with role = doctor AND clearance = high” can decrypt a medical query, or “LLMs with project_ID = 1 OR role = auditor” can access a certain log. Unlike RBAC, ABE does not grant blanket access to any LLM, while it evaluates the attributes presented at decryption time, enabling contextual decisions (attributes include an LLM’s role, affiliation, and user prompts).

Beyond these, emerging research on dynamic access control provides additional tools to strengthen zero-trust LLM systems. For example, Rezazadeh et al. (Rezazadeh et al., 2025) presented a “collaborative memory” framework for multi-LLM systems that encodes memory access permissions as a time-evolving bipartite graph linking users, LLMs, and resources. Specifically, the framework maintains separate private and shared memory tiers for each user, with each memory fragment tagged by provenance (which LLMs contributed it, which LLMs can access it, etc.). Context-aware and fine-grained read/write policies then operate on this graph: a read policy might say that LLM $A$ can retrieve data from user $B$’s shared memory only if an edge exists between $A$ and $B$ and the fragment is not time-expired, etc., resulting in filtered, transformed views of memory for each LLM.

5.3. Stateless and Ephemeral LLM Management

Traditionally, multiple LLMs maintain a shared memory pool to preserve states, which accumulate information across multiple requests to optimize inference performance and enable contextual understanding (Hu et al., 2024; Rezazadeh et al., 2025). However, this persistent state cannot be trusted, as malicious inputs can contaminate shared memory spaces, enabling cross-request information extraction and state poisoning attacks that compromise the integrity of future interactions. In contrast, stateless LLM management (Kwon et al., 2023; Prabhu et al., 2025) fundamentally rejects the traditional assumption that LLMs’ internal states can be trusted over time, instead requiring continuous verification of every component while minimizing the vulnerability window through short-lived, isolated execution contexts.

5.3.1. Stateless LLM

The evolution of stateless LLMs has been driven by breakthrough innovations in memory management and isolation techniques. PagedAttention (Kwon et al., 2023) established the foundational paradigm by implementing process-like isolation where each request is dynamically allocated with memory blocks. This approach partitions the key-value cache into fixed-size blocks with copy-on-write semantics, ensuring complete isolation between requests and automatic memory reclamation upon completion, thereby eliminating the persistent shared state. Building on this foundation, vAttention (Prabhu et al., 2025) advanced stateless security through hardware-level isolation using CUDA virtual memory APIs (CUD, [n. d.]), creating true memory virtualization where each request operates in protected virtual address spaces. BlockLLM (Arshad and Halim, 2025) extended stateless principles to multi-tenant environments through component-level micro-segmentation, dividing LLMs into cryptographically isolated blocks where each tenant processes requests through separate block instances. This architecture prevents cross-tenant information leakage while implementing per-block access control policies that ensure compromises cannot propagate beyond the current request, directly implementing zero-trust micro-segmentation principles at the finest possible granularity.

5.3.2. Ephemeral LLM

Ephemeral LLM management (Henderson et al., 2023) extends the stateless paradigm by introducing temporally-bounded LLM instances that automatically self-destruct after predetermined conditions are met, directly implementing zero-trust’s “assume breach” principle through proactive threat containment. Henderson et al. (Henderson et al., 2023) pioneered self-destructing models, which embed algorithmic time locks that degrade model performance when adapted for harmful tasks such as generating toxic content or bypassing safety guardrails, ensuring that compromised instances lose their utility for malicious purposes without external intervention. This approach transforms LLM instances into inherently untrusted, which aligns with zero-trust’s fundamental rejection of persistent trust relationships. Furthermore, ephemeral LLMs can be deployed on serverless computing platforms (Fu et al., 2023) to realize automatic lifecycle management with multiple security checkpoints.

6. Inter-LLM Zero-trust Security: System Approaches

In this section, we review the technical progress of system-level approaches, which ensure zero-trust security between LLMs and eventually build the zero-trust multi-LLM system.

6.1. Proactive Maintenance

Traditional perimeter-based security approaches typically employ reactive response mechanisms, taking actions after attacks are launched (Yu et al., 2025; Chen et al., 2025b; Kim et al., 2024). However, in multi-LLM collaborative environments, this reactive nature may allow malicious information to propagate and amplify throughout the edge network before detection. The zero-trust principle of “never trust, always verify” demands more proactive security maintenance strategies that can anticipate and prevent security incidents before they materialize.

6.1.1. Intelligent Input Checking

An effective proactive strategy involves input checking, specifically analyzing and verifying user prompts to preemptively detect and neutralize potential threats before the LLM generates harmful content. For instance, Kim et al. (Kim et al., 2024) proposed a comprehensive LLM prompt detection system that employs regular expressions and fine-tuned LLMs to identify personally identifiable information, malicious codes, URLs, and prompt injection commands within user prompts. Chen et al. (Chen et al., 2025a) introduced DefensiveTokens, which embed special tokens into user prompts. These tokens can disrupt or alter malicious patterns in the raw prompt, making it more difficult for attackers to construct successful injection attacks. Similarly, Zhang et al. (Zhang et al., 2025) presented JailGuard, which defends prompt-based attacks by creating multiple mutated variants of the user prompt (e.g., random replacement, targeted insertion). Muliarevych et al. (Muliarevych, 2024) explored a layered filtering framework, which includes a prompt analyzer to preprocess and wrap user inputs, and an attack validator that evaluates the wrapped prompts for malicious intent. Utilizing GPT-4 as the attacker validator, this layered filtering subsystem achieved superior effectiveness, accurately detecting 97% of various prompt injection attacks. SecurityLingua (Li et al., 2025) further enhanced proactive security by employing security-aware prompt compression. This method uses a specialized compressor model trained on malicious prompt patterns to effectively highlight critical security-related instructions within user prompts. By compressing the prompts, SecurityLingua removes extraneous, potentially distracting context that attackers often introduce to mask harmful intentions.

6.1.2. Reputation Schemes

In addition to input verification, the reputation of each LLM, which is based on its historical behavior and content generation quality, can be utilized as a metric to perform proactive maintenance. Specifically, LLMs whose reputation falls below a predefined threshold due to suspicious activities or poor-quality content can be proactively isolated or assigned lower priorities, thus minimizing potential damage. For instance, the LLMChain framework (Bouchiha et al., 2024) demonstrated such a reputation system, evaluating LLM reliability based on previous interactions, thereby identifying potentially malicious LLMs and mitigating the damage they can cause. In addition to reputation, other assessments can be applied from interdisciplinary perspectives. PsySafe (Zhang et al., 2024) presented a psychology-based framework that assesses multi-LLM systems by evaluating LLM behaviors and psychological states, allowing early identification of malicious tendencies and proactive intervention.

6.1.3. Topology-aware Proactive Maintenance

Beyond individual LLM assessments, network-wide safety evaluations play a crucial role in predictive security maintenance. MedSentry (Chen et al., 2025b), for instance, implemented a comprehensive evaluation pipeline that systematically benchmarks the risk existing in representative multi-LLM topologies. This allows vulnerabilities to be identified early, facilitating preemptive actions to mitigate misinformation or harmful content before impacting practical services. NetSafe (Yu et al., 2025) further extended this concept by examining the topological aspects of LLM interactions, highlighting how network configurations influence the propagation of misinformation or harmful behaviors, thus determining preventive strategies. Similarly, G-Safeguard (Wang et al., 2025) employed a temporal graph modeling technique to detect and remediate anomalies in LLM interactions. By analyzing communication patterns over time, G-Safeguard identified anomalous LLMs and disrupted the spread of harmful content, demonstrating significant effectiveness in containing and neutralizing threats early in their propagation. Lastly, Guardian (Zhou et al., 2025) leveraged topology-guided graph neural networks to proactively identify malicious LLMs within collaborative multi-LLM environments (see Fig. 5(left)). By systematically screening, adjudicating, and isolating malicious LLMs based on their interactions, Guardian maintains robust decision-making and prevents information contamination.

Lesson Learned: These three proactive maintenance approaches demonstrate a progressive defense strategy with increasing scope and sophistication. Input checking provides immediate threat prevention by analyzing individual prompts at the entry point, offering rapid responses. Reputation schemes advance this approach by leveraging historical behavioral data to proactively isolate potentially malicious LLMs based on accumulated trust scores, enabling predictive intervention before threats materialize. Network topology-aware maintenance operates from a macroscopic perspective, analyzing system-wide interaction patterns and communication flows to identify vulnerabilities and anomalies across the entire multi-LLM ecosystem.

6.2. Blockchain and Distributed Management

Blockchain is a decentralized ledger system that enables tamper-resistant record keeping on public networks (Liu et al., 2019; Huang et al., 2025; Lei et al., 2020). A distinctive characteristic of blockchain is the elimination of trusted authorities, operating under the assumption that any node could potentially be malicious, and achieving Byzantine fault tolerance through distributed data storage and consensus mechanisms (Leng et al., 2022). This property makes blockchain particularly valuable for constructing zero-trust multi-LLM environments in EGI scenarios. To be specific, first, distributed data storage preserves interaction records and historical content generated by LLMs across multiple nodes, preventing records from being tampered with by any single entity. Additionally, blockchain leverages consensus mechanisms to resist Byzantine attacks, where malicious nodes may attempt to manipulate the collaborative process or inject false information (Luo et al., 2025). Finally, smart contracts define various operational rules and trigger conditions that can be automatically executed without human intervention and cannot be influenced by attackers, making blockchains ideal for managing collaborative processes and responding to anomalous behaviors (Karanjai and Shi, 2024).

For data storage, Mo et al. (Mo et al., 2025) presented a blockchain-based crowdsourcing evaluation framework for LLMs, where LLMs’ reputation scores are recorded immutably on the blockchain. This approach ensures that evaluation scores cannot be retroactively modified, maintains a transparent history of LLM performance, and enables fair reputation-based selection of high-quality LLMs for future tasks. LLM-Net (Chong et al., 2025) constructed a multi-LLM network that divides LLMs into different roles, including coordinators, respondents, and validators. Validators execute blockchain consensus mechanisms to verify content generated by respondents, while coordinators run smart contracts to manage the collaborative process. Furthermore, all interactions and transactions are preserved on the blockchain, ensuring transparency and accountability. Similarly, Luo et al. (Luo et al., 2025) also utilized blockchain to construct decentralized multi-LLM networks. Notably, they evaluated the efficiency of four mainstream consensus algorithms in supporting multi-LLM operations and provided insights for selecting the most suitable one based on network latency and throughput requirements. BlockAgents (Chen et al., 2024) proposed a novel consensus mechanism, namely Proof-of-Thought (PoT), which is designed specifically for multi-LLM systems. Specifically, PoT ensures that LLMs contributing the most valuable insights to the collaborative reasoning process acquire the highest voting priority, while incorporating stake-based miner designation and multi-round debate-style voting to prevent malicious LLMs from dominating the decision-making process. Wang et al. (Wang et al., 2024) developed a threshold signature algorithm suitable for blockchain-assisted multi-LLM systems. Decisions agreed upon by multiple LLMs are collaboratively signed by everyone and written on the blockchain, enabling traceability and accountability while ensuring that no single entity can control the entire signature process. Karanjai et al. (Karanjai and Shi, 2024) employed smart contracts to manage the execution of LLM inference on edge devices, preventing attackers from launching DDoS attacks on EGI.

Apart from blockchain, Multi-Party Secure Computation (MPC) (Xu et al., 2025) has emerged as a critical technique for constructing zero-trust multi-LLM systems by enabling secure collaborative computations without revealing individual data inputs. This is particularly important in scenarios involving federated fine-tuning and decentralized inference, where maintaining data confidentiality and preventing unauthorized access to sensitive model information are paramount. Additionally, Zero-Knowledge Proofs (ZKP) (Qu et al., 2025) enhance zero-trust principles by enabling the verification of computations across multiple LLMs without revealing the underlying LLM parameters or private data. Qu et al. (Qu et al., 2025) introduced zkGPT, a non-interactive ZKP framework specifically tailored for efficient LLM inference. By generating cryptographic proofs that verify the correctness of inference results without exposing confidential model parameters, zkGPT aligns perfectly with the principles of zero-trust architectures, ensuring verifiability and integrity of computations in distributed multi-LLM systems.

6.3. Micro-segmentation and Isolation

Micro-segmentation (NIS, [n. d.]) is a critical zero-trust security technique that divides network infrastructure into smaller, isolated segments to limit lateral movement (i.e., attackers moving from one compromised system to adjacent systems within the network) and contain potential breaches. In EGI with multiple LLMs, micro-segmentation becomes particularly critical due to the diverse service requirements, sensitive data processing, and heterogeneous device configurations.

Recent advances in wireless network slicing have demonstrated significant potential for enhancing security isolation in multi-LLM deployments. Liu et al. (Liu et al., 2025) introduced WiLLM, the first open-source wireless communication system specifically designed for LLM services through its innovative “Tree-Branch-Fruit” slicing architecture. This hierarchical framework enables dedicated communication channels for different LLM services, ensuring that computational and communication resources are isolated between different slices while allowing telecom operators to monetize services through slice subscriptions. The approach directly supports zero-trust principles by establishing strict boundaries between LLM services, preventing unauthorized inter-LLM communication, and enabling granular access controls at the network level. Based on this foundation, Liu et al. (Liu et al., 2024) further demonstrated the practical implementation of dedicated network slicing for LLMs through their proposed LLM-Slice system, which creates LLM-specific network slices to efficiently bind services and communication resources. The system enables different LLMs, including Google Bard888https://gemini.google.com/app, Meta LLaMA999https://www.llama.com/, and ChatGPT (Brown et al., 2020), to coexist with independent resource allocation and management, ensuring that potential security breaches in one slice cannot propagate to others. Furthermore, a permissions database is used to enforce strict authentication policies for each LLM. Finally, an intelligent controller continuously validates cross-slice communications, enabling real-time threat detection that aligns with zero-trust’s continuous monitoring (Liu et al., 2024).

6.4. Intelligent System Monitoring and Failure Management

Zero-trust multi-LLM systems in EGI demand continuous monitoring and robust failure management to ensure reliability and safety. Recent studies have summarized numerous failure modes unique to multi-LLM setups that underscore the need for comprehensive monitoring (Joshi, 2025). For instance, Cemri et al. (Pan et al., 2025) identified 14 distinct failure modes in popular multi-LLM frameworks, ranging from coordination breakdowns (i.e., one LLM’s action invalidates another’s) to more subtle issues like shared misconceptions that propagate across the network. The complexity and variety of these failure modes highlight why traditional guardrails are insufficient for multi-LLM systems, necessitating more sophisticated monitoring approaches that can detect distributed failures.

Multi-LLM monitoring has followed a clear evolutionary path, progressively aligning with the stringent demands of zero-trust architectures. Initially, oversight relied on manually crafted rules or static guardrails, such as reactive output filters or step-by-step validators (Huang et al., 2024) to intercept disallowed actions. Despite effectively detecting straightforward violations, such as unauthorized tool calls or toxic outputs, these methods fundamentally fall short of zero-trust requirements due to their inability to dynamically adapt to evolving threats beyond predefined rules. To this end, researchers introduced traditional anomaly detection and machine learning techniques (Elhaminia et al., 2023), capitalizing on the rich telemetry from multi-LLM systems, such as action and interaction logs, performance metrics, and state-change traces, to identify deviations from expected behavior. Although these approaches expand monitoring capabilities and handle extensive data, they treat LLM systems as black-box output generators, lacking the nuanced semantic understanding essential for comprehensive oversight in zero-trust environments.

The latest advancement employs LLMs themselves as monitoring agents. For instance, Zhang et al. (Zhang et al., 2025) proposed AgentFM, which organizes multiple specialized LLMs to monitor activities specific to its designated role. Additionally, a higher-level meta-LLM integrates these role-specific insights, enabling the detection and handling of cross-role anomalies, effectively supporting the nuanced oversight required by zero-trust environments. As shown in Fig. 5(right), He et al. (He et al., 2025) introduced SentinelAgent, which implements real-time anomaly detection and intervention by dynamically constructing interaction graphs of LLM interactions. This framework detects anomalies at various granularities: individual LLM misbehavior at the node level, unsafe inter-LLM communications at the edge level, and suspicious exploit chains at the path level. Notably, a pluggable LLM-based monitoring agent is leveraged for semantic analysis, enabling active intervention such as interrupting harmful behaviors and correcting multi-LLM collusion, thereby offering robust zero-trust monitoring capabilities suited for complex multi-LLM ecosystems. Chang et al. (Chang and Geng, 2025) presented SagaLLM, which structures each task into independent, compensable units. In this framework, each operational LLM is paired with a dedicated compensation LLM responsible for rolling back and correcting any faulty operations. Independent validating LLMs are employed to assess intermediate and final outputs, detecting inconsistencies such as logical contradictions or coordination errors between LLMs. Upon detecting anomalies, SagaLLM automatically initiates structured rollbacks.

Remark 2.

The technical progress reviewed in Sections 5 and 6 contributes to the implementation and advancement of multi-LLM zero-trust systems in EGI (such as the one demonstrated in Section 4). For instance, the strong identity protocols and context-aware access control can realize the “Identity and Authentication Module”, proactive maintenance enables “User Input Checking” and “Multi-Layer LLM Output Verification” modules, and intelligent monitoring schemes realize “Behavioral Auditing and Anomaly Detection.”

7. Challenges and Future Research Directions

7.1. Ethical and Societal Issues

The deployment of zero-trust multi-LLM systems in EGI raises ethical and societal implications that extend beyond technical security considerations. Unlike traditional AI systems, multi-LLM systems in EGI directly interact with critical societal infrastructure, autonomous vehicles, and healthcare networks, amplifying the potential impact of algorithmic bias, discrimination, and social harm (Tiwari and Farag, 2025). The zero-trust paradigm’s fundamental assumption of “never trust, always verify” introduces additional ethical complexities when applied to human-AI interactions, potentially undermining user trust and creating psychological barriers to system adoption.

Future research should address several critical challenges: 1) developing ethical frameworks for algorithmic accountability in distributed multi-LLM decision-making, where responsibility attribution becomes complex due to the collaborative nature of reasoning processes and the absence of centralized control mechanisms, 2) establishing fairness-preserving zero-trust protocols that prevent discriminatory outcomes while maintaining security guarantees, particularly when LLMs trained on biased datasets collaborate in safety-critical applications, and 3) designing transparent governance mechanisms that enable public oversight and democratic participation in zero-trust multi-LLM system deployment decisions.

7.2. Asymmetric Information and Network Heterogeneity

The heterogeneous nature of EGI networks presents fundamental challenges in implementing unified zero-trust multi-LLM frameworks due to asymmetric information sharing capabilities and widely varying communication conditions among edge devices (Kherraf et al., 2019). The stark disparity between high-capacity wired backhaul connections (e.g., between roadside units and base stations with negligible delay) and unreliable wireless links (e.g., vehicle-to-vehicle channels with random delays and intermittent connectivity) creates temporal inconsistencies that can undermine zero-trust security protocols requiring synchronized verification and continuous authentication.

This network asymmetry introduces several critical research challenges: 1) developing delay-tolerant zero-trust protocols that can maintain security guarantees despite variable network conditions and intermittent connectivity, ensuring that authentication and authorization mechanisms remain effective even when some nodes experience communication delays or temporary isolation, 2) designing adaptive information sharing strategies that dynamically adjust collaboration patterns based on real-time network conditions while preserving the principle of least privilege access control, and 3) establishing distributed consensus mechanisms for multi-LLM coordination that are resilient to network partitions and asymmetric information propagation delays. Future work can also explore federated zero-trust architectures that leverage edge-cloud hierarchies to compensate for local network limitations, implementing tiered security policies that can gracefully degrade while maintaining critical safety properties when network conditions deteriorate.

7.3. Privacy-Preserving Collaborative Reasoning

The deployment of zero-trust multi-LLM systems necessitates collaborative reasoning capabilities that can leverage collective intelligence while maintaining cryptographic privacy guarantees and preventing any form of information leakage, extending the principle of “least privilege” to collaborative inference scenarios (Cao et al., 2025). This challenge is particularly acute in EGI scenarios where sensitive data must be processed at the edge while maintaining mathematically provable confidentiality.

Future research should address: 1) developing transformer-oriented encryption schemes that enable encrypted multi-LLM collaboration without any plaintext data exposure, ensuring that even compromised LLMs cannot access sensitive information (Cao et al., 2025), and 2) advancing secure MPC frameworks with ZKP that enable collaborative inference among distributed LLMs while maintaining cryptographic proof that no participating LLM can extract more information than strictly necessary for the collaborative task. Recent work on encryption-friendly LLMs demonstrates the feasibility of homomorphic encryption for transformer models, achieving significant computational speedups while maintaining performance comparable to plaintext models (Cao et al., 2025). However, extending these approaches to multi-LLM collaborative scenarios introduces additional complexity in terms of zero-trust encrypted communication protocols and cryptographically verifiable consensus mechanisms that must be addressed through innovative cryptographic and distributed systems research that assumes no trust in any component.

8. Conclusion

This paper has presented the first systematic survey of zero-trust security principles applied to multi-LLM systems in EGI. As multi-LLM systems become increasingly critical, traditional perimeter-based security approaches prove inadequate to address the unique vulnerabilities in collaborative edge deployments. We have systematically analyzed security challenges in multi-LLM systems, categorizing threats at both intra-LLM and inter-LLM levels, and demonstrate the limitations of existing trustworthy approaches. Then, we have proposed a unified zero-trust multi-LLM framework implementing the four fundamental zero-trust principles through detailed architectural design and operational workflows. We have categorized zero-trust mechanisms into model-level approaches focusing on individual LLM security and system-level approaches addressing distributed coordination challenges. Additionally, we have identified several critical future research directions that require immediate attention. We hope that this survey prompts both theoretical progress and practical implementation for secure edge agentic AI.