A Security-Oriented Lifecycle Model for Large Language Model Systems

Large language models (LLMs) are increasingly embedded in critical infrastructure, yet existing lifecycle frameworks are primarily designed for operational efficiency rather than security. This misalignment creates a significant gap: security-critical activities—such as data provenance, model signing, and agentic permission control—are often treated as implicit or secondary. Furthermore, current governance frameworks (like the NIST AI RMF or the EU AI Act) are not clearly mapped to specific lifecycle stages, leaving organizations without a clear roadmap for where to implement security controls.

The authors introduce a new lifecycle model comprising 32 distinct stages organized into four core pipeline layers: Data, Model, Distribution, and Application. By focusing on attack-surface boundaries—defined by the threat actors, artifacts, and access models present at each stage—the authors identify 13 stages that were previously merged or overlooked in standard MLOps frameworks. For example, the model distinguishes between Supervised Fine-Tuning (SFT) and Alignment, as these stages involve different threat actors and distinct poisoning vectors. This granular approach allows for more precise security analysis and risk management.

To support the core pipeline, the framework includes two cross-cutting pillars:

• LLMOps Pillar: A 12-stage operational support layer synthesized from academic and industry best practices to ensure consistent security and compliance automation.
• Governance Pillar: A 9-category mapping that links requirements from the NIST AI RMF, the EU AI Act, and ISO/IEC 42001 to specific lifecycle stages.

A key insight from this mapping is the structural asymmetry in the current regulatory landscape: governance evidence (documentation, logging) is concentrated at deployment-facing stages where regulators have visibility, while the most consequential security decisions (data selection, alignment strategy, capability boundaries) occur at development-facing stages where regulatory oversight is weakest.