SoK: A taxonomy of LLM threats

Securing large language models (LLMs) is critical as deployments scale, yet existing guidance is fragmented and phase-specific. This Systematization of Knowledge paper presents a unified, lifecycle-centric taxonomy of 20 attack classes spanning training, inference, and system integration. We synthesize academic and industry work into a dependency-aware threat schema that addresses causal and associative interdependencies, grounded in real-world case studies and visual mappings. Building on this, we develop a defense-in-depth strategy for coverage-oriented risk assessment, mitigation prioritization, security posture measurement, and open research directions in adversarial ML and secure systems engineering.