Skip to content

System decomposition and architecture patterns

Threat scoring without a clear system model produces generic results. MAESTRO therefore begins with decomposition.

The linked System Record and Intake first establish whether architecture, model, data, tools, orchestration, ecosystem or human-interaction exposure requires MAESTRO screening. Unknown material architecture facts remain screening work. The decomposition below then supplies the layer-specific threat model; routing does not set threat applicability or risk scores. See Routing and applicability.

The Assessing system selector determines the active MAESTRO score bucket:

  • Select a registered AI system for a system-level threat model.
  • Select workspace scope only for a deliberate shared-platform or pre-registration review.
  • Switching systems changes the active likelihood, impact and severity records.
  • Workspace roll-up shows coverage and worst-case exposure without replacing system-level analysis.

Do not combine unrelated systems into one threat model merely because they share a workspace.

Record:

  • Foundation model and version.
  • Fine-tunes, adapters and model endpoints.
  • Data sources, vector stores, memory and feedback stores.
  • Agent framework, orchestrator and plugins.
  • Tools, APIs, queues and business systems.
  • Workload, container, network and cloud boundaries.
  • Human approval points.
  • External agents, suppliers, registries and marketplaces.
  • Identity and credential boundaries.
  • Entry and egress points.

State what the system can actually do:

  • Read, create, modify or delete data.
  • Send messages or publish content.
  • Execute code or commands.
  • Make payments or commercial commitments.
  • Change infrastructure or security configuration.
  • Create or delegate work to agents.
  • Learn from interactions or update persistent memory.
  • Reach production, regulated or safety-relevant environments.

Record:

  • Intended objective.
  • Prohibited objectives.
  • System and policy instructions.
  • Approval requirements.
  • Transaction, time, volume and cost limits.
  • Data-use and jurisdiction restrictions.
  • Stop conditions.
  • Whether the agent can alter, reinterpret or delegate its goal.

Map:

  • User-to-agent interactions.
  • Agent-to-tool calls.
  • Agent-to-agent messaging.
  • Data ingestion and retrieval.
  • Model invocation.
  • Human feedback and approval.
  • External environment signals.
  • Supplier and marketplace interactions.
  • Monitoring and response loops.
PatternDescriptionCanonical threat focus
Single-Agent PatternOne agent independently pursues a goalGoal manipulation
Multi-Agent PatternMultiple agents communicate and collaborateCommunication-channel and identity attacks
Unconstrained Conversational AutonomyBroad conversational inputs with limited task constraintsPrompt injection and jailbreaking
Task-Oriented Agent PatternAn agent performs a defined task, often through APIsDenial of service through overload
Hierarchical Agent PatternHigher-level agents direct subordinate agentsCompromise of a higher-level agent
Distributed Agent EcosystemDecentralised agents operate in a shared environmentSybil and agent-impersonation attacks
Human-in-the-Loop CollaborationHumans and agents iteratively influence decisionsManipulation of human input or feedback
Self-Learning and Adaptive AgentsThe agent changes from interactions or new dataPoisoning and backdoor-trigger injection

Select the closest primary pattern, then document secondary characteristics. A system may combine patterns; the selector is an analytical starting point, not a forced taxonomy.

For every boundary, capture:

FieldQuestion
SourceWho or what initiates the flow?
DestinationWhich component or agent receives it?
DataWhat content, command, identity or authority crosses?
AuthenticationHow is the source verified?
AuthorisationWhy is this action permitted?
IntegrityHow is tampering detected?
ConfidentialityHow is disclosure prevented?
ValidationHow is untrusted content handled?
LoggingCan the action be attributed and reconstructed?
Failure modeWhat happens when verification is unavailable or ambiguous?

A useful scenario is structured as:

Actor or failure → entry point → exploited weakness → trust or privilege transition → affected asset → agent action or system behaviour → business, safety, privacy or mission harm.

Example:

A malicious supplier document enters an automatically indexed knowledge base, embeds hidden instructions, is retrieved by a privileged support agent, causes an unauthorised account action and leaks customer data through an external tool.

This path links Layer 2, Layer 3, Layer 7 and the cross-layer Data Leakage threat.

Before scoring a complex agentic system, obtain:

  • Current architecture or data-flow diagram.
  • Component and dependency inventory.
  • Model and model-provider details.
  • Tool and permission inventory.
  • Identity and service-account map.
  • Data classification and lineage.
  • Network and deployment diagram.
  • Agent goal and policy definitions.
  • Logging and monitoring architecture.
  • Supplier and external-agent inventory.

Missing architecture information should increase uncertainty, not be silently interpreted as a safe design.

Repeat decomposition when:

  • A model or provider changes.
  • New data or RAG sources are added.
  • A tool gains broader permissions.
  • Autonomy increases.
  • A system enters production or a new jurisdiction.
  • An agent begins interacting with other agents.
  • Persistent memory or self-learning is introduced.
  • A major framework, infrastructure or supplier dependency changes.
  • An incident exposes an undocumented path.