# TRiSM for Agentic AI: A Review of Trust, Risk, and Security Management in LLM-based Agentic Multi-Agent Systems

**Authors:** Shaina Raza, Ranjan Sapkota, Manoj Karkee, Christos Emmanouilidis  
**arXiv:** [2506.04133](https://arxiv.org/abs/2506.04133)  
**Date:** 2025-06-04  
**Summary page:** https://www.summarizepaper.com/en/paper/2506.04133v2/

## Key points

- Agentic AI systems are rapidly transforming intelligence, autonomy, collaboration, and decision-making in various sectors
- Trust, Risk, and Security Management (TRiSM) framework for Agentic Multi-Agent Systems (AMAS) powered by large language models (LLMs)
- Four pillars of TRiSM in Agentic AI: Governance, Explainability, ModelOps, Privacy/Security
- Introduction of a novel risk taxonomy to capture unique threats in Agentic AI like prompt injection and memory poisoning
- Proposal of new metrics - Component Synergy Score (CSS) and Tool Utilization Efficacy (TUE) - to assess inter-agent collaboration quality and tool efficiency
- Advocacy for advanced techniques like LIME and SHAP for enhancing explainability and robust security measures such as prompt hygiene and homomorphic encryption
- Research roadmap emphasizes scalable, verifiable, regulation-compliant Agentic AI systems focusing on advancing adversarial robustness and refining governance protocols
- Integration of TRiSM principles into the study of LLM-based multi-agent systems with a focus on critical domains like healthcare, science, and finance
- Prioritization of governance and security aspects as fundamental design considerations with actionable guidance for researchers addressing targeted threats like backdoor attacks in distributed LLMs
- These key points summarize the text's discussion on Trust, Risk, and Security Management within the context of Agentic AI systems powered by large language models.

## Summary

Agentic AI systems are rapidly transforming the landscape of intelligence, autonomy, collaboration, and decision-making in various sectors. This review focuses on Trust, Risk, and Security Management (TRiSM) within the context of Agentic Multi-Agent Systems (AMAS) powered by large language models (LLMs). It delves into the foundational concepts of Agentic AI and highlights its architectural differences from traditional AI agents. The framework for TRiSM in Agentic AI is structured around four pillars: Governance, Explainability, ModelOps, and Privacy/Security. These pillars are tailored to address the challenges specific to multi-agent LLM systems. A novel risk taxonomy is introduced to capture unique threats like prompt injection and memory poisoning in Agentic AI. Additionally, two new metrics - Component Synergy Score (CSS) and Tool Utilization Efficacy (TUE) - are proposed to assess inter-agent collaboration quality and tool efficiency. The review also discusses advanced techniques such as LIME and SHAP for enhancing explainability and advocates for robust security measures like prompt hygiene and homomorphic encryption to mitigate vulnerabilities. Looking ahead, the research roadmap emphasizes scalable, verifiable, and regulation-compliant Agentic AI systems. Future work should focus on advancing adversarial robustness, refining governance protocols, and establishing standardized benchmarks for trustworthiness and coordination. By aligning Agentic AI development with TRiSM principles , researchers can ensure responsible integration of these innovative systems into society. This review stands out by integrating TRiSM principles into the study of LLM-based multi-agent systems with a focus on critical domains like healthcare , science ,and finance . Unlike previous studies that overlook governance and security aspects , this work prioritizes these elements as fundamental design considerations. It provides actionable guidance for researchers while addressing targeted threats like backdoor attacks in distributed LLMs and governance challenges related to regulatory compliance across autonomous agents. The literature review methodology followed best practices in systematic reviews to ensure thorough coverage and rigor. Key research objectives included identifying trustworthiness challenges in Agentic AI systems enabled by LLMs; exploring technical strategies to address these challenges; mapping existing approaches onto TRiSM pillars; identifying gaps in current research; and outlining promising directions for future work. Overall trends show a significant increase in Agentic AI research post-2023 with a growing focus on governance and safety topics within the TRiSM framework. This comprehensive review fills critical gaps in both technical multi-agent literature and broader responsible AI landscape while providing valuable insights for researchers, engineers, policymakers involved in developing trustworthy Agentic AI systems.