Prompt Injection Attacks in Large Language Models and AI Agent Systems: A Comprehensive Review of Vulnerabilities, Attack Vectors, and Defense Mechanisms
Атаки с внедрением промптов на большие языковые модели и системы ИИ-агентов: всесторонний обзор уязвимостей, векторов атак и механизмов защиты
2026-01-07
SCID: 54.1/6nemwvgh
Discuss with AI
AI agent systemsModel Context ProtocolRAG poisoninglarge language modelsprompt injection attacks
Figures from the paper
Abstract (AI)
Large Language Models (LLMs) have rapidly transformed artificial intelligence applica-tions across industries, yet their integration into production systems has unveiled critical security vulnerabilities, chief among them prompt injection attacks. This comprehensive review synthesizes research from 2023-2025, analyzing over 120 peer-reviewed papers, industry security reports, and documented real-world exploits. We examine the taxonomy of prompt injection techniques, including direct jailbreaking and indirect injection through external content. The rise of AI agent systems and the Model Context Protocol (MCP) has dramatically expanded attack surfaces, introducing vulnerabilities such as tool poisoning and credential theft. We document critical incidents including GitHub Copilot's CVE-2025-53773 remote code execution vulnerability (CVSS 9.6) and ChatGPT's Windows license key exposure. Research demonstrates that just five carefully crafted documents can manipulate AI responses 90% of the time through Retrieval-Augmented Generation (RAG) poisoning. This review provides actionable mitigation strategies based on OWASP Top 10 for LLM Applications 2025, identifies fundamental limitations including the stochastic nature problem and alignment paradox, and proposes research directions for architectur-ally secure AI systems. Our analysis reveals that prompt injection represents a funda-mental architectural vulnerability requiring defense-in-depth approaches rather than singular solutions.
Key Findings
1
AI agents and the Model Context Protocol expand attack surfaces through vulnerabilities including tool poisoning and credential theft.
2
Documented incidents include GitHub Copilot’s CVE-2025-53773 remote code execution vulnerability, rated CVSS 9.6, and ChatGPT Windows license key exposure.
3
Only five carefully crafted documents can manipulate AI responses 90% of the time through Retrieval-Augmented Generation poisoning.
4
Prompt injection is identified as a fundamental architectural vulnerability requiring defense-in-depth, exemplified by the proposed five-layer PALADIN framework, rather than a single mitigation.
5
The review synthesizes 45 sources from 2023–2025, classifying prompt injection into direct jailbreaking and indirect attacks through external content.
Research Object
large language models and AI agent systems integrated into production applications
Research Subject
prompt injection vulnerabilities, attack vectors, impacts, and defense mechanisms, including architectural security limitations
Publication Details
Publication Date
2026-01-07
Journal
Publisher
ISSN
Cited by
45
Open access PDF
Access Type
Author Information
Download PDF
Subscribe to digest