On Understanding, Identifying, and Mitigating Vulnerabilities in Agentic Large Language Models

2026-08-11

AI Summary

This study addresses the critical security gap in large language model (LLM)-based autonomous agents endowed with real-world operational capabilities, where reasoning components remain highly vulnerable to attacks that can trigger unauthorized access, irreversible state changes, or cascading failures. Conducting a systematic literature review of 85 studies from 2023–2025 following PRISMA 2020 guidelines, this work reveals a pronounced imbalance between attack and defense research (3.9:1) and introduces the first four-layer vulnerability taxonomy for agent LLMs—encompassing perception, cognition, action, and interaction—identifying 13 vulnerability types and seven open challenges centered on isolation. Notably, perception-layer vulnerabilities account for 66% of reported issues, whereas high-risk action-layer threats such as tool misuse and sandbox escape constitute only 4.7%, highlighting a severe misalignment between research focus and actual risk. The study further identifies cross-layer vulnerability propagation due to architectural coupling as the root cause of systemic security weaknesses.

Abstract

Large Language Models (LLMs) have undergone a shift from stateless conversational interfaces to autonomous agents capable of multi-step planning, tool invocation, code execution, and maintaining persistent memory. When these agents operate with real-world privileges---calling APIs, modifying files, and querying databases---a compromised reasoning step can trigger unauthorized data access, irreversible state changes, or cascading failures, yet the security research community has not kept pace. To quantify the state of the field, we conducted a systematic literature review under PRISMA 2020 guidelines across six databases, screening 743 records and retaining 85 papers (2023--2025) on agentic LLM security. Attack research outpaces defense work by 3.9:1. Perception-layer vulnerabilities (prompt injection, jailbreaking, adversarial perturbations) dominate, accounting for 66\% of papers, while action-layer vulnerabilities (tool misuse, code injection, sandbox escape) appear in only 4.7\%, misaligned with real-world risk. Code execution security accounts for 3.5\%, and tool-augmented agents 12\%. We contribute a four-layer taxonomy mapping 13 vulnerability types across perception, brain, action, and interaction layers, and identify seven open problems centered on containment. Agentic LLM insecurity stems from architectural coupling, where weak isolation allows vulnerabilities to propagate across layers.

Problem

• Agentic LLMs
• security vulnerabilities
• action-layer vulnerabilities
• architectural coupling
• containment

Innovation

• agentic LLMs
• vulnerability taxonomy
• architectural coupling
• security isolation
• systematic literature review