LLM Security in Cloud-Native Architectures: A Comprehensive Survey of Attacks, Defenses, and Operational Challenges

The rapid adoption of Large Language Models (LLMs) in cloud-native environments has introduced unprecedented security challenges that traditional cybersecurity frameworks fail to adequately address. This survey provides a comprehensive analysis of LLM security threats in cloud infrastructures, examining 65 recent studies and real-world incidents from 2023-2025. We propose a novel taxonomy that categorizes threats into four primary dimensions: Infrastructure Security (credential theft, resource abuse), Application Security (prompt injection, data exfiltration), Model Security (model poisoning, adversarial attacks), and Supply Chain Security (dependency vulnerabilities, malicious packages). Our analysis reveals that LLMjacking attacks have increased 10x in 2024, with single incidents costing victims up to $100,000 per day. We identify critical research gaps, including the lack of standardized security frameworks for LLM-integrated applications and insufficient real-time threat detection capabilities. This survey synthesizes current defensive strategies, evaluates their effectiveness, and proposes future research directions toward achieving secure and trustworthy LLM deployments in cloud environments. Our findings provide actionable insights for security architects, cloud engineers, and researchers working to establish robust security postures for AI-powered cloud applications.