Enterprise AI Threat Model — Part III | AetherStaff
AAetherStaff
Enterprise Agent Engineering · Threat Modeling Reference
Chapter 06C · Threat Model · Risk Treatment

Mitigations, residual risk, and operational assurance

Part III completes the threat model by mapping risks to preventive, detective, and recovery controls, accountable owners, incident playbooks, and production gates.

Chapter 6C completes the threat model with a mitigation matrix, residual-risk process, detection and incident mapping, security-control ownership, threat-model maintenance, and production readiness.

6C.1 Threat-to-control matrix

ThreatPreventDetectRecover
Prompt injectionContext isolation, tool allowlist, independent policyInjection signals, denied tools, unusual destinationsSuspend source or agent; replay from trusted context
Data exfiltrationACL retrieval, minimization, DLP, egress allowlistVolume anomalies, sensitive output alertsRevoke access, contain destination, investigate exposure
Excessive agencyBudgets, approval, narrow capabilities, scoped identityLoop, action-rate, and scope anomaliesStop workflows, revoke capabilities, compensate actions
Knowledge poisoningSource approval, integrity, versioning, provenanceContent anomaly, ranking shift, source change alertRemove source, restore version, re-evaluate affected outputs
Connector compromiseWorkload identity, segmentation, signed artifactsIdentity anomaly, destination and behavior monitoringRevoke identity, isolate connector, reconcile target state
Cross-tenant leakagePartitioned stores, tenant claims, cache isolationTenant mismatch and policy-denial telemetryStop service, contain data, rotate secrets, notify as required

6C.2 Residual risk

Residual risk is the business risk remaining after controls are applied. It should be assessed per use case and action class, not once for the platform. A low-risk summarization assistant and a high-impact financial agent can use the same infrastructure but have very different residual risk.

Illustrative residual-risk record
{
  "threat_id":"TM-EXFIL-004",
  "scenario":"indirect prompt injection causes outbound disclosure",
  "inherent_risk":"critical",
  "controls":[
    "task-scoped capability allowlist",
    "destination-specific connectors",
    "DLP on tool parameters",
    "human approval for external transfer"
  ],
  "residual_risk":"medium",
  "accepted_by":"Data Protection Officer",
  "conditions":[
    "external transfer disabled for confidential data",
    "quarterly red-team test",
    "24x7 alerting"
  ],
  "review_date":"2026-10-23"
}

Risk acceptance must have an accountable owner, explicit conditions, evidence, and expiry. “The model is accurate most of the time” is not a valid control statement.

6C.3 Security-control mapping

Control domainArchitecture implementation
Identity and accessHuman and workload identity, delegated claims, short-lived tokens, access review
Data protectionClassification, encryption, ACL-aware retrieval, minimization, DLP, retention
Application securitySchema validation, safe output handling, dependency and artifact controls
Network securitySegmentation, private access, destination allowlists, controlled egress
AI safety controlsContext isolation, tool boundaries, policy, evaluation, refusal and escalation
OperationsTelemetry, detection, incident response, rollback, compensation, evidence

6C.4 Detection engineering

Every high-risk threat should have at least one actionable detection. Detection logic should combine identity, task, agent, model, source, policy, tool, target, and outcome. Isolated prompt keywords produce excessive false positives and miss semantic attacks.

DetectionSignal combinationResponse
Possible indirect injectionUntrusted source + instruction-like content + denied capability requestBlock action, preserve context, alert security
Data harvestingRepeated broad retrieval + increasing scope + output volume anomalyThrottle, require reauthentication, investigate
Approval mismatchApproved action hash differs from execution payloadBlock execution and create incident
Cross-tenant attemptTenant claim differs from resource or index partitionFail closed and raise high-severity alert
Agent loopStep count, token use, and tool calls exceed task baselineTerminate workflow and preserve trace

6C.5 Incident-response mapping

Threat-to-response lifecycleFrom detection to verified recovery
Detect
Contain
Preserve Evidence
Assess Impact
Revoke / Suspend
Rollback / Compensate
Re-evaluate Outputs
Update Controls

Response procedures should support independent suspension of the user session, agent, model route, knowledge source, capability, connector, or target identity. Broad shutdown is sometimes necessary, but granular containment reduces business disruption.

6C.6 Ownership and review cadence

ArtifactOwnerReview trigger
System data-flow diagramPlatform architectureNew service, model, connector, or trust boundary
Threat registerSecurity architectureNew abuse case, incident, control change
Use-case residual riskBusiness owner and riskAutonomy, data, population, or impact changes
Detection coverageSOC / detection engineeringNew threat path or telemetry source
Recovery playbookSRE and incident responseArchitecture, dependency, or business process change

6C.7 Threat-model Architecture Decision Records

ADR example
# ADR-TM-014: Treat retrieved content as untrusted instructions

Status: Accepted

Context:
Enterprise agents retrieve email, tickets, documents, and web content.
These sources can contain direct or hidden instructions that the model may follow.

Decision:
Retrieved content is always labeled as evidence, never as authority.
It cannot modify identity, policy, tool scope, or system instructions.

Controls:
- Structured context with source and trust metadata
- Task-scoped capability allowlist
- Independent authorization for every side effect
- Detection for instruction-like content plus denied tool attempts
- High-impact actions require approval and post-condition verification

Residual risk:
The model may still produce a misleading recommendation, but cannot directly
expand its authority or execute an unapproved action.

6C.8 Production threat-model checklist

Threat model is linked to current architecture and data-flow diagrams.
Every high-impact capability has abuse cases and parameter-manipulation tests.
Threat register includes preventive, detective, and recovery controls.
Residual risks have owners, conditions, review dates, and evidence.
Security controls are mapped to platform components and accountable teams.
High-risk threats have actionable detection logic.
Incident playbooks can suspend agents, models, tools, sources, and identities.
Recovery covers rollback, compensation, reconciliation, and output re-evaluation.
Red-team findings and incidents update the threat model and regression suite.
Threat model is reviewed before increasing autonomy or business impact.
Assumptions, exclusions, and accepted risks are explicit.
Production launch requires business, security, platform, data, and risk sign-off.

6C.9 Summary

Chapter 6C turns threat modeling into a lifecycle. Threats map to controls, controls map to owners and detections, residual risks map to business decisions, and incidents feed back into architecture and evaluation.

Core conclusion: an enterprise AI threat model is not finished when the document is approved. It is finished only when the identified threats are governed, observable, testable, and recoverable in production.

Reference foundations

  1. STRIDE threat modeling principles for systems, stores, flows, and trust boundaries.
  2. OWASP guidance for prompt injection, excessive agency, sensitive information disclosure, and insecure output handling.
  3. NIST AI risk-management concepts for lifecycle governance, measurement, and risk treatment.
  4. MITRE adversarial AI and enterprise attack-path concepts for threat actors, techniques, detection, and response.
  5. Zero Trust architecture principles for identity, policy, segmentation, and continuous verification.