A Four-Layer Security Governance Framework for LLM-Based AI Agents

作者：Yiang Gao, Shanshan Wu · 发表于：Journal of Artificial Intelligence Practice · 年份：2025 · DOI：10.23977/jaip.2025.080406 · 被引用次数：2

: As artificial intelligence advances from “dialogue intelligence” to “decision intelligence,” AI agents built upon Large Language Models (LLMs) are becoming a crucial force driving transformation across industries. However, their autonomous capabilities in perception, decision-making, memory, and execution introduce systemic security risks far beyond traditional LLM vulnerabilities. This paper presents a four-layer security governance framework covering the full Perception–Decision–Memory– Execution lifecycle to mitigate risks such as multi-source perception failures, decision hallucination, memory poisoning, and malicious execution. By systematically mapping each lifecycle phase to security requirements and controls, this framework provides theoretically grounded and practically applicable guidance for the trustworthy and secure development of AI agents.