// AGENTIC-AI THREAT MODELING

MAESTRO: Seven-Layer Reference Model for Agentic AI

Decompose the architecture layer-by-layer, then map threats per layer. Layer 6 is cross-cutting: it spans all layers.
LAYER07
Agent Ecosystem
Multi-agent interaction · agent marketplaces · external services & APIs
LAYER05
Evaluation & Observability
Monitoring · logging · eval harnesses · drift & anomaly detection
LAYER04
Deployment & Infrastructure
Runtime environments · containers · compute, network & scaling
LAYER03
Agent Frameworks
Agent orchestration · tool integration · reasoning & planning loops
LAYER02
Data Operations
RAG pipelines · vector stores · data ingestion & provenance
LAYER01
Foundation Models
Core LLMs & generative models · weights · fine-tuning surface
06
Security & Compliance
CSA / Ken Huang. The seven layers give a shared decomposition: use MAESTRO to break the architecture apart, ATLAS to enumerate adversarial-ML techniques per layer, and the AI Controls Matrix to select controls. Layer 6 (Security & Compliance) is drawn as a vertical band because it is not one slice: it runs through every layer.
© 2026 Ken Connell. All rights reserved.
AGENTIC SECOPS