Skip to main content

AI Threat Model

Traditional applications execute predefined code paths. AI agents also interpret natural language, retrieve untrusted content, select tools, and generate actions. The security boundary therefore includes every source that can influence the model and every system the model can affect.

Attack Surface

Review these surfaces for every production agent:

SurfaceTypical riskPrimary control
User promptInstruction override, data extraction, unsafe requestInput policy, Prompt Shield, scoped permissions
Retrieved document or websiteIndirect prompt injection, poisoned knowledgeTreat content as data, source review, output constraints
System prompt and skillsExcessive authority or unclear boundariesControlled editing, version review, least privilege
Memory and data collectionsUnauthorized or stale contextOwnership, classification, access filters, review cadence
Model outputHallucination, harmful or biased resultEvaluation, grounding, human review, content safety
Tool responseMalicious instructions or unexpected dataValidate tool output, isolate it from instructions
Tool executionExternal communication, deletion, spend, privilege changeFunction-level policy and human approval
Model providerProcessing location, retention, availability, model changeApproved model inventory and deployment review

Minimum Threat-Modelling Questions

For each agent or workflow, document:

  1. Who can invoke it and whether anonymous access is possible.
  2. Which datasets, memory collections, files, and page context it can read.
  3. Which tool functions it can call and which functions modify external state.
  4. What a malicious user or malicious document could cause it to disclose or change.
  5. Which actions require human approval and who can approve them.
  6. Which events are logged and how an investigator can correlate them.
  7. Which model deployment processes prompts and where processing may occur.
  8. What happens when a provider, credential, index, or approval flow fails.

Review Trigger

Repeat the threat model when access, tools, datasets, prompts, models, public interfaces, or workflow logic changes. A previously low-risk assistant can become high risk as soon as it receives a write-capable connection or sensitive data source.

See Prompt Injection, Tool Governance, and Layered Security Architecture.